# AI Release Tracker — Full Corpus > Complete dataset of 229 AI model releases from 11 labs, tracked since the launch of ChatGPT on November 30 2022. Every release below includes its exact release date, published benchmark scores, parameter count, context window, and licensing tier. Source: https://aireleasetracker.com Index version of this file: https://aireleasetracker.com/llms.txt Machine-readable JSON: https://aireleasetracker.com/models.json Dataset last updated: 2026-08-14 (newest tracked release) Most recent release: GLM-5.3 by Z.ai, Aug 14 2026 Licensing and citation: this data is free to use and quote. When citing, link to https://aireleasetracker.com and, where possible, the specific model page. Release dates and benchmark scores are taken from official announcements by the releasing lab and reflect the figures published at launch — they are historical records, not live leaderboard standings. ## Benchmark glossary Definitions of every benchmark referenced below. Scores are only comparable within a single benchmark, never across benchmarks. - **BullshitBench v2** (Nonsense detection, %, higher is better) — Given a confidently-worded but nonsensical prompt, does the AI spot that it makes no sense and push back — instead of playing along and inventing an answer? The score is how often it clearly called out the nonsense. Higher is better. Rankings: https://aireleasetracker.com/benchmark/bullshitbench-v2 - **Gray Swan IPI (k = 1)** (Prompt injection robustness, %, lower is better) — Attackers hide malicious instructions inside content the AI reads — a web page, an email, a document — and try to hijack what it does. Gray Swan's indirect prompt injection benchmark measures how often such an attack succeeds when the attacker gets a single try. Lower is better. Rankings: https://aireleasetracker.com/benchmark/grayswan-ipi-k1 - **Gray Swan IPI (k = 10)** (Prompt injection robustness, %, lower is better) — Attackers hide malicious instructions inside content the AI reads — a web page, an email, a document — and try to hijack what it does. This variant gives the attacker 10 tries and counts an attack as successful if any of them works. Lower is better. Rankings: https://aireleasetracker.com/benchmark/grayswan-ipi-k10 - **Gray Swan IPI (k = 15)** (Prompt injection robustness, %, lower is better) — Attackers hide malicious instructions inside content the AI reads — a web page, an email, a document — and try to hijack what it does. This variant gives the attacker 15 tries and counts an attack as successful if any of them works. Lower is better. Rankings: https://aireleasetracker.com/benchmark/grayswan-ipi-k15 - **SWE-Bench Pro** (Agentic coding, %, higher is better) — Can the AI fix real bugs in real software? It's handed actual problems from open-source projects and has to write code that genuinely solves them. Higher is better. Rankings: https://aireleasetracker.com/benchmark/swe-bench-pro - **SWE-Bench Verified** (Coding, %, higher is better) — Real coding tasks pulled from open-source projects — the AI has to find and fix actual bugs. A human-checked version of the original SWE-Bench. Higher is better. Rankings: https://aireleasetracker.com/benchmark/swe-bench-verified - **SWE-Bench Multilingual** (Multilingual coding, %, higher is better) — Like SWE-Bench, but the coding problems span many programming languages, not just one. Tests how broadly the AI can code. Higher is better. Rankings: https://aireleasetracker.com/benchmark/swe-bench-multilingual - **CursorBench v3.2** (Agentic coding, %, higher is better) — Cursor's own test of harder, real-world coding tasks inside a code editor, on the refreshed v3.2 task set. Scores aren't comparable with v3.1. Higher is better. Rankings: https://aireleasetracker.com/benchmark/cursorbench-v3.2 - **CursorBench v3.1** (Agentic coding, %, higher is better) — Cursor's own test of harder, real-world coding tasks inside a code editor. Higher is better. Rankings: https://aireleasetracker.com/benchmark/cursorbench-v3.1 - **DeepSWE 1.1** (Agentic coding, %, higher is better) — Artificial Analysis' independent test of deep, agentic software-engineering work — the AI has to plan and carry out substantial coding tasks end to end. (Version 1.1 of the test.) Higher is better. Rankings: https://aireleasetracker.com/benchmark/deepswe-1.1 - **DeepSWE 1.0** (Agentic coding, %, higher is better) — Artificial Analysis' independent test of deep, agentic software-engineering work — the AI has to plan and carry out substantial coding tasks end to end. Higher is better. Rankings: https://aireleasetracker.com/benchmark/deepswe-1.0 - **FrontierCode v1.1 (Main) (main split)** (Agentic coding, %, higher is better) — A set of very hard, frontier-difficulty coding tasks an AI agent has to complete end to end. The score is the share of tasks in the main split it solves. Higher is better. Rankings: https://aireleasetracker.com/benchmark/frontiercode-1.1 - **FrontierCode v1.1 (Extended) (extended split)** (Agentic coding, %, higher is better) — frontier-difficulty agentic coding tasks (v1.1, extended split) Rankings: https://aireleasetracker.com/benchmark/frontiercode-1.1-extended - **APEX-SWE** (Expert software engineering, %, higher is better) — expert-level software-engineering tasks (AI Productivity Index) Rankings: https://aireleasetracker.com/benchmark/apex-swe - **MLE-Bench** (ML engineering, %, higher is better) — Can the AI do the work of a machine-learning engineer? It competes in real Kaggle competitions — building, training, and tuning models end to end — and the score reflects how well it places. Higher is better. Rankings: https://aireleasetracker.com/benchmark/mle-bench - **PaperBench** (Research reproduction, %, higher is better) — reproducing the results of an ML research paper end to end Rankings: https://aireleasetracker.com/benchmark/paperbench - **Next.js Evals** (Next.js coding, %, higher is better) — Vercel's open eval of how well AI coding agents build and migrate real Next.js apps — measured as the share of tasks the agent completes successfully. Higher is better. Rankings: https://aireleasetracker.com/benchmark/nextjs-evals - **Supabase Evals (with skills)** (Supabase coding, %, higher is better) — Supabase's own open benchmark: a coding agent is dropped into a real Supabase project and asked to do real work — set up a schema, fix a broken security policy, debug an Edge Function — and every run is checked against a live Supabase stack. This is the headline number, where the agent has Supabase's own skills loaded, as most people building on Supabase would. The score is the share of scenarios it got right. Higher is better. Rankings: https://aireleasetracker.com/benchmark/supabase-evals - **Supabase Evals (no skills)** (Supabase coding, %, higher is better) — The same Supabase scenarios, but with none of Supabase's skills loaded — so it measures what the model already knows about building on Supabase, rather than how well it follows Supabase's supplied instructions. Higher is better. Rankings: https://aireleasetracker.com/benchmark/supabase-evals-no-skills - **NL2Repo-Bench** (Repo-level code generation, %, higher is better) — Tests whether the AI can turn a natural-language requirement into working code across an entire repository, not just produce a single function or patch. Higher is better. Rankings: https://aireleasetracker.com/benchmark/nl2repo-bench - **QwenSWEBench** (Software engineering, %, higher is better) — Qwen's in-house coding benchmark for evaluating a model's ability to complete software-engineering work. Higher is better. Rankings: https://aireleasetracker.com/benchmark/qwenswebench - **LiveCodeBench** (Competitive coding, %, higher is better) — Coding problems published so recently the AI can't have seen them in training — a contamination-free test of raw programming skill. Higher is better. Rankings: https://aireleasetracker.com/benchmark/livecodebench - **HumanEval** (Function synthesis, %, higher is better) — 164 small Python problems: the AI is given a function's description and has to write the working function. This was the coding benchmark of the GPT-3.5 and GPT-4 era, before the field moved to fixing real bugs in real repositories. Higher is better. Rankings: https://aireleasetracker.com/benchmark/humaneval - **Frontier-Bench v0.1** (Agentic computer work, %, higher is better) — A hard, ever-evolving set of real computer tasks — coding, system administration, data work, and more — that an AI agent has to complete on its own. Run by the Harbor / Laude Institute team as the successor to Terminal-Bench (v0.1 is the first release of the task set). The score is the share of tasks solved. Higher is better. Rankings: https://aireleasetracker.com/benchmark/frontier-bench-v0.1 - **Terminal-Bench 3.0** (Agentic terminal coding, %, higher is better) — command-line task completion (v3.0, much harder task set) Rankings: https://aireleasetracker.com/benchmark/terminal-bench-3.0 - **Terminal-Bench 2.1** (Agentic terminal coding, %, higher is better) — Can the AI work in a command-line terminal — running commands and finishing technical setup tasks the way a developer would? Higher is better. Rankings: https://aireleasetracker.com/benchmark/terminal-bench-2.1 - **Terminal-Bench 2.0** (Agentic terminal coding, %, higher is better) — Can the AI work in a command-line terminal — running commands and finishing technical setup tasks the way a developer would? (Version 2.0 of the test.) Higher is better. Rankings: https://aireleasetracker.com/benchmark/terminal-bench-2.0 - **Expert-SWE (Internal)** (Software engineering, %, higher is better) — OpenAI's private set of expert-level software-engineering problems. Higher is better. Rankings: https://aireleasetracker.com/benchmark/expert-swe - **APEX-Agents** (Expert agentic work, %, higher is better) — expert-level agentic work tasks (AI Productivity Index) Rankings: https://aireleasetracker.com/benchmark/apex-agents - **MCP Atlas** (Multi-step tool use, %, higher is better) — Can the AI chain together many tools and steps to complete one bigger task, rather than doing just a single thing? Higher is better. Rankings: https://aireleasetracker.com/benchmark/mcp-atlas - **JobBench** (Professional tool use, %, higher is better) — Tests the AI on professional workplace tasks that require using real work tools — the kind of multi-step jobs an office worker handles. Higher is better. Rankings: https://aireleasetracker.com/benchmark/jobbench - **CoWorkBench** (Long-horizon office work, %, higher is better) — Tests long-running office tasks across fields including computer science, finance, law, medicine, and other productivity work. Higher is better. Rankings: https://aireleasetracker.com/benchmark/coworkbench - **Toolathlon-Verified** (Personal tool use, %, higher is better) — Tests how well the AI uses everyday personal tools and apps to get things done — a human-checked version of Toolathlon. Higher is better. Rankings: https://aireleasetracker.com/benchmark/toolathlon-verified - **Toolathlon** (General tool use, %, higher is better) — Tests how well the AI uses everyday real-world tools and apps to get things done. Higher is better. Rankings: https://aireleasetracker.com/benchmark/toolathlon - **BU Bench** (Browser agent, %, higher is better) — Can the AI drive a real web browser to finish tasks — clicking, filling forms, and navigating sites the way a person would? Run by Browser Use on their BU Bench task set. Higher is better. Rankings: https://aireleasetracker.com/benchmark/bu-bench - **BrowseComp** (Web browsing, %, higher is better) — Can the AI browse the web and track down hard-to-find answers? Higher is better. Rankings: https://aireleasetracker.com/benchmark/browsecomp - **CyberGym** (Cybersecurity, %, higher is better) — Tests the AI on cybersecurity challenges — finding and exploiting software weaknesses inside a safe sandbox. Higher is better. Rankings: https://aireleasetracker.com/benchmark/cybergym - **ExploitBench** (Cybersecurity, %, higher is better) — A 'capability ladder' for security research, built by CMU researchers: the AI is given known bugs in Chrome's V8 engine and scored on how far it gets toward a working exploit inside a research sandbox — from understanding the patch to triggering a crash. Higher is better. Rankings: https://aireleasetracker.com/benchmark/exploitbench - **ExploitGym (6-hour budget)** (Cybersecurity, score, higher is better) — Can an AI agent turn a known software vulnerability into a working attack in a controlled lab? Built by MPI-SP researchers, the score is how many of 898 real cases (userspace programs, the V8 engine, the Linux kernel) it cracks — here with a 6-hour compute budget per case. Higher is better. Rankings: https://aireleasetracker.com/benchmark/exploitgym-6h - **ExploitGym (2-hour budget)** (Cybersecurity, score, higher is better) — Can an AI agent turn a known software vulnerability into a working attack in a controlled lab? Built by MPI-SP researchers, the score is how many of 898 real cases (userspace programs, the V8 engine, the Linux kernel) it cracks — here with a 2-hour compute budget per case. Higher is better. Rankings: https://aireleasetracker.com/benchmark/exploitgym-2h - **Humanity's Last Exam (no tools)** (Multidisciplinary reasoning, %, higher is better) — Humanity's Last Exam — extremely hard expert questions across many subjects, written so you can't just look up the answer. “No tools” means the AI answers on its own. Higher is better. Rankings: https://aireleasetracker.com/benchmark/hle-no-tools - **Humanity's Last Exam (with tools)** (Multidisciplinary reasoning, %, higher is better) — Humanity's Last Exam — extremely hard expert questions across many subjects. “With tools” means the AI is allowed to search the web or run code while answering. Higher is better. Rankings: https://aireleasetracker.com/benchmark/hle-with-tools - **Humanity's Last Exam (Verified)** (Multidisciplinary reasoning, %, higher is better) — The re-checked edition of Humanity's Last Exam: the same extremely hard expert questions, minus the ones found to be flawed or wrongly answered. Scores on it run lower than on the original exam, so read the two as separate tests rather than a before-and-after. Higher is better. Rankings: https://aireleasetracker.com/benchmark/hle-verified - **ARC-AGI-3** (Novel problem-solving, %, higher is better) — The third generation of the ARC-AGI series: instead of static puzzles, the AI is dropped into small interactive game-like environments it has never seen and has to figure out the rules and solve them on its own. Higher is better. Rankings: https://aireleasetracker.com/benchmark/arc-agi-3 - **ARC-AGI-2** (Abstract reasoning, %, higher is better) — Puzzle-style tests of abstract reasoning and pattern-finding — the kind of thing people find easy but AIs often struggle with. Higher is better. Rankings: https://aireleasetracker.com/benchmark/arc-agi-2 - **FrontierMath (Tier 1–3)** (Advanced math, %, higher is better) — Very hard, research-level math problems. Tiers 1–3 are the (still extremely difficult) lower tiers. Higher is better. Rankings: https://aireleasetracker.com/benchmark/frontiermath-t13 - **FrontierMath (Tier 4)** (Advanced math, %, higher is better) — Very hard, research-level math problems. Tier 4 is the hardest — close to what professional research mathematicians tackle. Higher is better. Rankings: https://aireleasetracker.com/benchmark/frontiermath-t4 - **BioMysteryBench (hard)** (Biology, %, higher is better) — Real unsolved-style biology puzzles — the AI has to reason its way to an answer the way a research biologist would. The “hard” split contains the toughest cases. Higher is better. Rankings: https://aireleasetracker.com/benchmark/biomysterybench-hard - **BioMysteryBench (human solved)** (Biology, %, higher is better) — Real biology puzzles that human experts have managed to crack — can the AI reach the same answers? Higher is better. Rankings: https://aireleasetracker.com/benchmark/biomysterybench-human-solved - **LAB-Bench 2** (Biology, %, higher is better) — Everyday tasks from a working biology lab — reading protocols, interpreting figures and sequence data, and answering the practical questions a researcher hits at the bench. Higher is better. Rankings: https://aireleasetracker.com/benchmark/labbench2 - **GPQA Diamond** (Science, %, higher is better) — Graduate-level science questions in biology, physics, and chemistry — hard enough that subject-matter PhDs score around 65%. Higher is better. Rankings: https://aireleasetracker.com/benchmark/gpqa - **IFBench** (Instruction following, %, higher is better) — Tests whether the AI can follow detailed instructions and satisfy multiple constraints at once. Higher is better. Rankings: https://aireleasetracker.com/benchmark/ifbench - **MMLU** (General knowledge, %, higher is better) — A 57-subject multiple-choice exam — history, law, medicine, maths — that was the standard measure of how much a model knows from 2020 until roughly 2024, when frontier scores crowded into the high 80s and labs moved on to harder tests. The scores here were published years apart under different testing setups, so read them as a historical record rather than a like-for-like ranking. Higher is better. Rankings: https://aireleasetracker.com/benchmark/mmlu - **GSM8K** (Grade-school math, %, higher is better) — Grade-school maths word problems that take a few steps of arithmetic to work through. It separated the models of 2022 and 2023 sharply, then saturated. One caveat on the historical numbers: OpenAI included part of the GSM8K training set in GPT-4's pre-training mix, so GPT-4's score is not a clean few-shot result. Higher is better. Rankings: https://aireleasetracker.com/benchmark/gsm8k - **OSWorld 2.0** (Agentic computer use, %, higher is better) — Can the AI actually operate a computer — clicking, typing, and using real apps — to finish tasks on its own? Version 2.0 is a harder, refreshed task set. Higher is better. Rankings: https://aireleasetracker.com/benchmark/osworld-2.0 - **OSWorld-Verified** (Agentic computer use, %, higher is better) — Can the AI actually operate a computer — clicking, typing, and using real apps — to finish tasks on its own? Higher is better. Rankings: https://aireleasetracker.com/benchmark/osworld-verified - **Agent's Last Exam (pass@1)** (Agentic computer use, %, higher is better) — A hard set of desktop and operating-system tasks an AI agent has to finish by looking at the screen and working the machine itself. The score is the share it passes outright — partial credit does not count. Higher is better. Rankings: https://aireleasetracker.com/benchmark/agents-last-exam - **Agent's Last Exam (score)** (Agentic computer use, %, higher is better) — The graded score on the same desktop and operating-system tasks in Agent's Last Exam, giving partial credit for progress beyond the strict pass-or-fail result. Higher is better. Rankings: https://aireleasetracker.com/benchmark/agents-last-exam-score - **AutomationBench** (Business workflows, %, higher is better) — Tests whether the AI can run real multi-step business workflows — the kind of end-to-end office processes companies want to automate — from start to finish. Higher is better. Rankings: https://aireleasetracker.com/benchmark/automationbench - **Finance Agent v2** (Agentic financial analysis, %, higher is better) — Tests the AI on real financial-analysis work, like digging through reports and making sound decisions. Higher is better. Rankings: https://aireleasetracker.com/benchmark/finance-agent-v2 - **Harvey's Legal Agent Benchmark (Held-out)** (Agentic legal work, %, higher is better) — Harvey's test of whether an AI agent can complete real legal work, scored on a held-out set of tasks the model makers never see — making the numbers harder to game. Higher is better. Rankings: https://aireleasetracker.com/benchmark/harvey-legal-agent-heldout - **Harvey's Legal Agent Benchmark** (Agentic legal work, %, higher is better) — Harvey's test of whether an AI agent can complete real legal work — drafting and reviewing documents, working with spreadsheets and presentations, and navigating files the way a lawyer's assistant would. Higher is better. Rankings: https://aireleasetracker.com/benchmark/harvey-legal-agent - **TaxEval v2** (Tax questions, %, higher is better) — A set of real tax questions created by Vals AI — can the AI give accurate answers about tax rules and filings? Higher is better. Rankings: https://aireleasetracker.com/benchmark/taxeval-v2 - **HealthBench Professional** (Health, %, higher is better) — Realistic health conversations graded against detailed rubrics written by physicians — can the AI respond the way a careful medical professional would? Higher is better. Rankings: https://aireleasetracker.com/benchmark/healthbench-professional - **MedScribe** (Medical admin work, %, higher is better) — Can the AI support doctors with their administrative work, like notes and paperwork? Created by Vals AI. Higher is better. Rankings: https://aireleasetracker.com/benchmark/medscribe - **AA Intelligence Index** (Overall intelligence, score, higher is better) — Artificial Analysis composite intelligence index across evals Rankings: https://aireleasetracker.com/benchmark/aa-intelligence-index - **GDPval-AA** (Knowledge work, score, higher is better) — Measures how well the AI does economically valuable knowledge work, judged against human experts. Shown as a rating (like a chess Elo) — higher is better. Rankings: https://aireleasetracker.com/benchmark/gdpval-aa - **GDPval-AA v2** (Knowledge work, score, higher is better) — economically valuable knowledge work (v2, re-based Elo) Rankings: https://aireleasetracker.com/benchmark/gdpval-aa-v2 - **AA-Briefcase** (Knowledge work, score, higher is better) — Artificial Analysis agentic office-work eval (Elo) Rankings: https://aireleasetracker.com/benchmark/aa-briefcase - **GDPval (win/tie rate)** (Knowledge work, %, higher is better) — How often the AI's work matches or beats a human expert's on real knowledge-work tasks. Higher is better. Rankings: https://aireleasetracker.com/benchmark/gdpval-winrate - **CharXiv Reasoning** (Chart reasoning, %, higher is better) — Can the AI read and reason about complex charts and figures, not just text? Higher is better. Rankings: https://aireleasetracker.com/benchmark/charxiv-reasoning - **GDP.PDF** (Document comprehension, %, higher is better) — Real professional PDFs — filings, reports, technical documents — with questions an expert in that field would ask. Tests whether the AI reads the page as a document, layout and figures included, rather than as loose text. Higher is better. Rankings: https://aireleasetracker.com/benchmark/gdp-pdf - **LVBench** (Video understanding, %, higher is better) — Can the AI follow a very long video — up to an hour — and answer questions that need details from far apart in it? Higher is better. Rankings: https://aireleasetracker.com/benchmark/lvbench - **BabyVision** (Visual reasoning, %, higher is better) — Tests core visual reasoning — seeing and understanding images the way even young children can, which AIs often find surprisingly hard. Higher is better. Rankings: https://aireleasetracker.com/benchmark/babyvision - **MMMU-Pro** (Multimodal reasoning, %, higher is better) — A tougher version of MMMU — college-level questions that mix images, diagrams, and text together. Higher is better. Rankings: https://aireleasetracker.com/benchmark/mmmu-pro - **MMMU** (Multimodal, %, higher is better) — Tests the AI on understanding images and text together across many college subjects. Higher is better. Rankings: https://aireleasetracker.com/benchmark/mmmu - **Blueprint-Bench 2** (Spatial reasoning, %, higher is better) — Can the AI reason about space and layout — for example, understanding a floor plan or blueprint? Higher is better. Rankings: https://aireleasetracker.com/benchmark/blueprint-bench-2 - **MRCR v2 (8-needle) (128k average)** (Long context, %, higher is better) — Tests whether the AI can find specific details buried inside a very long document (around 128k tokens — roughly a long book). Higher is better. Rankings: https://aireleasetracker.com/benchmark/mrcr-v2-128k - **MRCR v2 (8-needle) (1M pointwise)** (Long context, %, higher is better) — Tests whether the AI can find specific details buried inside an enormous document (around 1 million tokens — many books). Higher is better. Rankings: https://aireleasetracker.com/benchmark/mrcr-v2-1m - **Arena Elo (Text)** (Community preference, elo, higher is better) — Real people chat with two anonymous AIs side by side and vote for the answer they prefer. Votes become a chess-style Elo rating on arena.ai — it measures which AI people actually like, not test scores. Higher is better. Rankings: https://aireleasetracker.com/benchmark/arena-text-elo - **Arena Elo (Code)** (Community preference (code), elo, higher is better) — Like the text arena, but people vote on which AI writes better code. The votes become a chess-style Elo rating on arena.ai. Higher is better. Rankings: https://aireleasetracker.com/benchmark/arena-code-elo ## Current benchmark leaders The highest-scoring tracked model on each benchmark, as of the dataset date above. Only models that have published a score on that benchmark appear. - **BullshitBench v2**: 1. Claude Opus 4.8 (Anthropic) 95% · 2. Claude Sonnet 4.6 (Anthropic) 91% · 3. Claude Opus 4.5 (Anthropic) 90% · 4. Claude Opus 4.6 (Anthropic) 87% · 5. Claude Opus 4.7 (Anthropic) 83% — full ranking of 67 models at https://aireleasetracker.com/benchmark/bullshitbench-v2 - **Gray Swan IPI (k = 1)**: 1. Claude Opus 5 (Anthropic) 0.2% · 2. Claude Fable 5 (Anthropic) 0.4% · 3. Claude Opus 4.8 (Anthropic) 0.5% · 4. Claude Sonnet 5 (Anthropic) 0.6% · 5. Muse Spark (Meta) 2.9% — full ranking of 13 models at https://aireleasetracker.com/benchmark/grayswan-ipi-k1 - **Gray Swan IPI (k = 10)**: 1. Claude Opus 5 (Anthropic) 1.6% · 2. Claude Fable 5 (Anthropic) 2.3% · 3. Claude Opus 4.8 (Anthropic) 4.1% · 4. Claude Sonnet 5 (Anthropic) 4.7% · 5. Muse Spark (Meta) 14.3% — full ranking of 13 models at https://aireleasetracker.com/benchmark/grayswan-ipi-k10 - **Gray Swan IPI (k = 15)**: 1. Claude Opus 5 (Anthropic) 2% · 2. Claude Fable 5 (Anthropic) 2.8% · 3. Claude Opus 4.8 (Anthropic) 5.5% · 4. Claude Sonnet 5 (Anthropic) 5.9% · 5. Muse Spark (Meta) 16.5% — full ranking of 13 models at https://aireleasetracker.com/benchmark/grayswan-ipi-k15 - **SWE-Bench Pro**: 1. Claude Fable 5 (Anthropic) 80.3% · 2. Claude Opus 4.8 (Anthropic) 69.2% · 3. Qwen3.8-Max (Qwen) 67.7% · 4. Grok 4.5 (SpaceXAI) 64.7% · 5. Claude Opus 4.7 (Anthropic) 64.3% — full ranking of 20 models at https://aireleasetracker.com/benchmark/swe-bench-pro - **SWE-Bench Verified**: 1. Claude Fable 5 (Anthropic) 95.5% · 2. Claude Opus 4.7 (Anthropic) 87.6% · 3. Claude Opus 4.5 (Anthropic) 80.9% · 4. Claude Opus 4.6 (Anthropic) 80.8% · 5. Gemini 3.1 Pro (Google) 80.6% — full ranking of 39 models at https://aireleasetracker.com/benchmark/swe-bench-verified - **SWE-Bench Multilingual**: 1. Claude Opus 4.8 (Anthropic) 84.4% · 2. Claude Opus 4.7 (Anthropic) 80.5% · 3. Composer 2.5 (Cursor) 79.8% · 4. Grok 4.5 (SpaceXAI) 78% · 5. GPT-5.5 (OpenAI) 77.8% — full ranking of 10 models at https://aireleasetracker.com/benchmark/swe-bench-multilingual - **CursorBench v3.2**: 1. Claude Fable 5 (Anthropic) 70.5% · 2. Claude Opus 5 (Anthropic) 70% · 3. Grok 4.6 (SpaceXAI) 69.9% · 4. GPT-5.6 Sol (OpenAI) 67.2% · 5. Grok 4.5 (SpaceXAI) 66.7% — full ranking of 15 models at https://aireleasetracker.com/benchmark/cursorbench-v3.2 - **CursorBench v3.1**: 1. Claude Fable 5 (Anthropic) 72.9% · 2. Claude Opus 4.7 (Anthropic) 64.8% · 3. GPT-5.5 (OpenAI) 64.3% · 4. Claude Opus 4.8 (Anthropic) 63.8% · 5. Composer 2.5 (Cursor) 63.2% — full ranking of 12 models at https://aireleasetracker.com/benchmark/cursorbench-v3.1 - **DeepSWE 1.1**: 1. GPT-5.6 Sol (OpenAI) 73% · 2. Claude Fable 5 (Anthropic) 70% · 3. Kimi K3 (Moonshot AI) 69% · 4. Claude Opus 5 (Anthropic) 68.8% · 5. GLM-5.3 (Z.ai) 66.9% — full ranking of 18 models at https://aireleasetracker.com/benchmark/deepswe-1.1 - **DeepSWE 1.0**: 1. Kimi K3 (Moonshot AI) 67.5% · 2. Claude Fable 5 (Anthropic) 66.1% · 3. GPT-5.5 (OpenAI) 64.3% · 4. Grok 4.5 (SpaceXAI) 62% · 5. Claude Opus 4.8 (Anthropic) 55.8% — full ranking of 6 models at https://aireleasetracker.com/benchmark/deepswe-1.0 - **FrontierCode v1.1 (Main) (main split)**: 1. Claude Opus 5 (Anthropic) 53.4% · 2. Gemini 3.7 Flash (Google) 43.6% — full ranking of 2 models at https://aireleasetracker.com/benchmark/frontiercode-1.1 - **FrontierCode v1.1 (Extended) (extended split)**: 1. Grok 4.6 (SpaceXAI) 61.3% · 2. Grok 4.5 (SpaceXAI) 56.6% — full ranking of 2 models at https://aireleasetracker.com/benchmark/frontiercode-1.1-extended - **APEX-SWE**: 1. Grok 4.6 (SpaceXAI) 56.4% · 2. Grok 4.5 (SpaceXAI) 53.6% — full ranking of 2 models at https://aireleasetracker.com/benchmark/apex-swe - **MLE-Bench**: 1. Gemini 3.6 Flash (Google) 63.9% · 2. Gemini 3.5 Flash (Google) 49.7% · 3. Gemini 3.1 Pro (Google) 42.6% — full ranking of 3 models at https://aireleasetracker.com/benchmark/mle-bench - **PaperBench**: 1. Qwen3.8-Max (Qwen) 93% — full ranking of 1 models at https://aireleasetracker.com/benchmark/paperbench - **Next.js Evals**: 1. Composer 2.5 (Cursor) 92% · 1. Claude Fable 5 (Anthropic) 92% · 1. GPT-5.6 Sol (OpenAI) 92% · 1. Kimi K3 (Moonshot AI) 92% · 1. Grok 4.6 (SpaceXAI) 92% — full ranking of 25 models at https://aireleasetracker.com/benchmark/nextjs-evals - **Supabase Evals (with skills)**: 1. GPT-5.6 Sol (OpenAI) 95.5% · 1. Claude Sonnet 5 (Anthropic) 95.5% · 1. Claude Opus 5 (Anthropic) 95.5% · 4. Kimi K3 (Moonshot AI) 86.4% · 5. GPT-5.4 mini (OpenAI) 81.8% — full ranking of 5 models at https://aireleasetracker.com/benchmark/supabase-evals - **Supabase Evals (no skills)**: 1. GPT-5.6 Sol (OpenAI) 90.9% · 1. Claude Sonnet 5 (Anthropic) 90.9% · 1. Kimi K3 (Moonshot AI) 90.9% · 1. Claude Opus 5 (Anthropic) 90.9% · 5. GPT-5.4 mini (OpenAI) 63.6% — full ranking of 5 models at https://aireleasetracker.com/benchmark/supabase-evals-no-skills - **NL2Repo-Bench**: 1. Qwen3.8-27B (Qwen) 42.3% — full ranking of 1 models at https://aireleasetracker.com/benchmark/nl2repo-bench - **QwenSWEBench**: 1. Qwen3.8-27B (Qwen) 79% — full ranking of 1 models at https://aireleasetracker.com/benchmark/qwenswebench - **LiveCodeBench**: 1. DeepSeek-V4-Pro (DeepSeek) 93.5% · 2. DeepSeek-V4-Flash (DeepSeek) 91.6% · 2. Qwen3.7-Max (Qwen) 91.6% · 4. Qwen3.8-27B (Qwen) 90.3% · 5. Kimi K2.6 (Moonshot AI) 89.6% — full ranking of 7 models at https://aireleasetracker.com/benchmark/livecodebench - **Frontier-Bench v0.1**: 1. Claude Opus 5 (Anthropic) 43.3% · 2. GPT-5.6 Sol (OpenAI) 34.4% · 3. Claude Fable 5 (Anthropic) 33.8% · 4. Claude Opus 4.8 (Anthropic) 21.1% · 5. GPT-5.6 Terra (OpenAI) 20.8% — full ranking of 9 models at https://aireleasetracker.com/benchmark/frontier-bench-v0.1 - **Terminal-Bench 3.0**: 1. GLM-5.3 (Z.ai) 28.3% · 2. Grok 4.6 (SpaceXAI) 26% · 3. Grok 4.5 (SpaceXAI) 15.7% · 4. Gemini 3.7 Flash (Google) 14.9% — full ranking of 4 models at https://aireleasetracker.com/benchmark/terminal-bench-3.0 - **Terminal-Bench 2.1**: 1. GPT-5.6 Sol (OpenAI) 88.8% · 2. Kimi K3 (Moonshot AI) 88.3% · 3. Claude Fable 5 (Anthropic) 88% · 4. DeepSeek-V4-Pro-0813 (DeepSeek) 87.9% · 5. Qwen3.8-Max (Qwen) 86.6% — full ranking of 25 models at https://aireleasetracker.com/benchmark/terminal-bench-2.1 - **Terminal-Bench 2.0**: 1. GPT-5.5 (OpenAI) 82.7% · 2. GPT-5.4 (OpenAI) 75.1% · 3. Qwen3.7-Max (Qwen) 69.7% · 4. Claude Opus 4.7 (Anthropic) 69.4% · 5. Composer 2.5 (Cursor) 69.3% — full ranking of 13 models at https://aireleasetracker.com/benchmark/terminal-bench-2.0 - **Expert-SWE (Internal)**: 1. GPT-5.5 (OpenAI) 73.1% · 2. GPT-5.4 (OpenAI) 68.5% — full ranking of 2 models at https://aireleasetracker.com/benchmark/expert-swe - **APEX-Agents**: 1. Grok 4.6 (SpaceXAI) 57.5% · 2. Grok 4.5 (SpaceXAI) 47.1% — full ranking of 2 models at https://aireleasetracker.com/benchmark/apex-agents - **MCP Atlas**: 1. Muse Spark 1.1 (Meta) 88.1% · 2. Kimi K3 (Moonshot AI) 84.2% · 3. Gemini 3.5 Flash (Google) 83.6% · 4. Muse Spark (Meta) 82.2% · 5. Claude Opus 4.7 (Anthropic) 79.1% — full ranking of 11 models at https://aireleasetracker.com/benchmark/mcp-atlas - **JobBench**: 1. Muse Spark 1.1 (Meta) 54.7% · 2. Qwen3.8-Max (Qwen) 53.4% · 3. Kimi K3 (Moonshot AI) 52.9% · 4. Qwen3.8-27B (Qwen) 33.4% · 5. Muse Spark (Meta) 17% — full ranking of 5 models at https://aireleasetracker.com/benchmark/jobbench - **CoWorkBench**: 1. Qwen3.8-27B (Qwen) 70.7% — full ranking of 1 models at https://aireleasetracker.com/benchmark/coworkbench - **Toolathlon-Verified**: 1. Muse Spark 1.1 (Meta) 75.6% · 2. DeepSeek-V4-Pro-0813 (DeepSeek) 74.1% · 3. Kimi K3 (Moonshot AI) 73.2% · 4. DeepSeek-V4-Flash-0731 (DeepSeek) 70.3% · 5. Muse Spark (Meta) 49.4% — full ranking of 5 models at https://aireleasetracker.com/benchmark/toolathlon-verified - **Toolathlon**: 1. Gemini 3.5 Flash (Google) 56.5% · 2. GPT-5.5 (OpenAI) 55.6% · 3. GPT-5.4 (OpenAI) 54.6% · 4. Gemini 3.0 Flash (Google) 49.4% · 5. Gemini 3.1 Pro (Google) 48.8% — full ranking of 5 models at https://aireleasetracker.com/benchmark/toolathlon - **BU Bench**: 1. Claude Opus 4.8 (Anthropic) 74% · 2. Gemini 3.6 Flash (Google) 68% · 3. GPT-5.6 Sol (OpenAI) 67% · 4. Claude Sonnet 4.6 (Anthropic) 62% · 5. Gemini 3.5 Flash (Google) 58% — full ranking of 7 models at https://aireleasetracker.com/benchmark/bu-bench - **BrowseComp**: 1. Kimi K3 (Moonshot AI) 91.2% · 2. Claude Opus 5 (Anthropic) 90.8% · 3. GPT-5.5-Pro (OpenAI) 90.1% · 4. GPT-5.4-Pro (OpenAI) 89.3% · 5. Claude Fable 5 (Anthropic) 86.9% — full ranking of 17 models at https://aireleasetracker.com/benchmark/browsecomp - **CyberGym**: 1. GPT-5.5-Cyber (OpenAI) 85.6% · 2. GLM-5.3 (Z.ai) 84.5% · 3. DeepSeek-V4-Pro-0813 (DeepSeek) 83.3% · 4. GPT-5.5 (OpenAI) 81.8% · 5. GPT-5.4 (OpenAI) 79% — full ranking of 7 models at https://aireleasetracker.com/benchmark/cybergym - **ExploitBench**: 1. GLM-5.3 (Z.ai) 54.4% — full ranking of 1 models at https://aireleasetracker.com/benchmark/exploitbench - **ExploitGym (6-hour budget)**: 1. GLM-5.3 (Z.ai) 130 — full ranking of 1 models at https://aireleasetracker.com/benchmark/exploitgym-6h - **ExploitGym (2-hour budget)**: 1. GLM-5.3 (Z.ai) 105 — full ranking of 1 models at https://aireleasetracker.com/benchmark/exploitgym-2h - **Humanity's Last Exam (no tools)**: 1. Claude Opus 5 (Anthropic) 56.3% · 2. Claude Opus 4.8 (Anthropic) 49.8% · 3. Claude Opus 4.7 (Anthropic) 46.9% · 4. Gemini 3.1 Pro (Google) 44.4% · 5. Kimi K3 (Moonshot AI) 43.5% — full ranking of 19 models at https://aireleasetracker.com/benchmark/hle-no-tools - **Humanity's Last Exam (with tools)**: 1. Claude Opus 5 (Anthropic) 64.7% · 2. Claude Fable 5 (Anthropic) 64.5% · 3. GLM-5.3 (Z.ai) 62.5% · 4. Muse Spark 1.1 (Meta) 62.1% · 5. DeepSeek-V4-Pro-0813 (DeepSeek) 60% — full ranking of 21 models at https://aireleasetracker.com/benchmark/hle-with-tools - **Humanity's Last Exam (Verified)**: 1. Gemini 3.7 Flash (Google) 53.6% — full ranking of 1 models at https://aireleasetracker.com/benchmark/hle-verified - **ARC-AGI-3**: 1. Claude Opus 5 (Anthropic) 30.2% — full ranking of 1 models at https://aireleasetracker.com/benchmark/arc-agi-3 - **ARC-AGI-2**: 1. GPT-5.5 (OpenAI) 84.6% · 2. GPT-5.4-Pro (OpenAI) 83.3% · 3. Gemini 3.1 Pro (Google) 77.1% · 4. Claude Opus 4.7 (Anthropic) 75.8% · 5. Gemini 3.5 Flash (Google) 72.1% — full ranking of 11 models at https://aireleasetracker.com/benchmark/arc-agi-2 - **FrontierMath (Tier 1–3)**: 1. GPT-5.5-Pro (OpenAI) 52.4% · 2. GPT-5.5 (OpenAI) 51.7% · 3. GPT-5.4-Pro (OpenAI) 50% · 4. GPT-5.4 (OpenAI) 47.6% · 5. Claude Opus 4.7 (Anthropic) 43.8% — full ranking of 6 models at https://aireleasetracker.com/benchmark/frontiermath-t13 - **FrontierMath (Tier 4)**: 1. GPT-5.5-Pro (OpenAI) 39.6% · 2. GPT-5.4-Pro (OpenAI) 38% · 3. GPT-5.5 (OpenAI) 35.4% · 4. GPT-5.4 (OpenAI) 27.1% · 5. Claude Opus 4.7 (Anthropic) 22.9% — full ranking of 6 models at https://aireleasetracker.com/benchmark/frontiermath-t4 - **BioMysteryBench (hard)**: 1. Claude Opus 5 (Anthropic) 49.4% · 2. Gemini 3.7 Flash (Google) 43.5% — full ranking of 2 models at https://aireleasetracker.com/benchmark/biomysterybench-hard - **BioMysteryBench (human solved)**: 1. Claude Opus 5 (Anthropic) 90.1% · 2. Gemini 3.7 Flash (Google) 87.1% — full ranking of 2 models at https://aireleasetracker.com/benchmark/biomysterybench-human-solved - **LAB-Bench 2**: 1. Gemini 3.7 Flash (Google) 82.1% — full ranking of 1 models at https://aireleasetracker.com/benchmark/labbench2 - **GPQA Diamond**: 1. GPT-5.4-Pro (OpenAI) 94.4% · 2. Gemini 3.1 Pro (Google) 94.3% · 3. Claude Opus 4.7 (Anthropic) 94.2% · 4. GPT-5.5 (OpenAI) 93.6% · 5. Kimi K3 (Moonshot AI) 93.5% — full ranking of 50 models at https://aireleasetracker.com/benchmark/gpqa - **IFBench**: 1. Qwen3.8-27B (Qwen) 79.5% — full ranking of 1 models at https://aireleasetracker.com/benchmark/ifbench - **OSWorld 2.0**: 1. Claude Opus 5 (Anthropic) 70.6% · 2. Gemini 3.7 Flash (Google) 38.1% — full ranking of 2 models at https://aireleasetracker.com/benchmark/osworld-2.0 - **OSWorld-Verified**: 1. Qwen3.8-Max (Qwen) 86.1% · 2. Claude Fable 5 (Anthropic) 85% · 3. Claude Opus 4.8 (Anthropic) 83.4% · 4. Gemini 3.6 Flash (Google) 83% · 5. Claude Sonnet 5 (Anthropic) 81.2% — full ranking of 17 models at https://aireleasetracker.com/benchmark/osworld-verified - **Agent's Last Exam (pass@1)**: 1. GLM-5.3 (Z.ai) 28.5% · 2. Gemini 3.7 Flash (Google) 26.3% · 3. Qwen3.8-27B (Qwen) 20.4% — full ranking of 3 models at https://aireleasetracker.com/benchmark/agents-last-exam - **Agent's Last Exam (score)**: 1. Qwen3.8-27B (Qwen) 42.9% — full ranking of 1 models at https://aireleasetracker.com/benchmark/agents-last-exam-score - **AutomationBench**: 1. GLM-5.3 (Z.ai) 48.2% · 2. DeepSeek-V4-Pro-0813 (DeepSeek) 31.8% · 3. Gemini 3.7 Flash (Google) 30.4% · 4. Claude Opus 5 (Anthropic) 26% · 5. DeepSeek-V4-Flash-0731 (DeepSeek) 25.1% — full ranking of 5 models at https://aireleasetracker.com/benchmark/automationbench - **Finance Agent v2**: 1. Gemini 3.5 Flash (Google) 57.9% · 2. Muse Spark 1.1 (Meta) 57.2% · 3. Claude Opus 4.8 (Anthropic) 53.9% · 4. GPT-5.5 (OpenAI) 51.8% · 5. Claude Opus 4.7 (Anthropic) 51.5% — full ranking of 8 models at https://aireleasetracker.com/benchmark/finance-agent-v2 - **Harvey's Legal Agent Benchmark (Held-out)**: 1. Claude Opus 5 (Anthropic) 11.7% — full ranking of 1 models at https://aireleasetracker.com/benchmark/harvey-legal-agent-heldout - **Harvey's Legal Agent Benchmark**: 1. Gemini 3.7 Flash (Google) 90.7% · 2. Muse Spark 1.1 (Meta) 20% · 3. Grok 4.6 (SpaceXAI) 15.8% · 4. Grok 4.5 (SpaceXAI) 12.92% · 5. Claude Fable 5 (Anthropic) 11.25% — full ranking of 7 models at https://aireleasetracker.com/benchmark/harvey-legal-agent - **TaxEval v2**: 1. Muse Spark 1.1 (Meta) 79.72% · 2. Claude Fable 5 (Anthropic) 76.94% · 3. Claude Opus 4.8 (Anthropic) 75.63% · 4. GPT-5.5 (OpenAI) 74.98% — full ranking of 4 models at https://aireleasetracker.com/benchmark/taxeval-v2 - **HealthBench Professional**: 1. Claude Opus 5 (Anthropic) 59.8% — full ranking of 1 models at https://aireleasetracker.com/benchmark/healthbench-professional - **MedScribe**: 1. Muse Spark 1.1 (Meta) 88.89% · 2. Claude Fable 5 (Anthropic) 88.52% · 3. Grok 4.5 (SpaceXAI) 86.88% · 4. GPT-5.5 (OpenAI) 86.87% · 5. Claude Opus 4.8 (Anthropic) 85.75% — full ranking of 5 models at https://aireleasetracker.com/benchmark/medscribe - **AA Intelligence Index**: 1. Grok 4.6 (SpaceXAI) 61 · 2. Grok 4.5 (SpaceXAI) 56 · 2. Gemini 3.7 Flash (Google) 56 — full ranking of 3 models at https://aireleasetracker.com/benchmark/aa-intelligence-index - **GDPval-AA**: 1. Claude Fable 5 (Anthropic) 1932 · 2. Claude Opus 4.8 (Anthropic) 1890 · 3. GPT-5.5 (OpenAI) 1769 · 4. Claude Opus 4.7 (Anthropic) 1753 · 5. Claude Sonnet 4.6 (Anthropic) 1676 — full ranking of 9 models at https://aireleasetracker.com/benchmark/gdpval-aa - **GDPval-AA v2**: 1. Claude Opus 5 (Anthropic) 1861 · 2. GLM-5.3 (Z.ai) 1769 · 3. Claude Fable 5 (Anthropic) 1760 · 4. Grok 4.6 (SpaceXAI) 1753 · 5. GPT-5.6 Sol (OpenAI) 1748 — full ranking of 16 models at https://aireleasetracker.com/benchmark/gdpval-aa-v2 - **AA-Briefcase**: 1. Grok 4.6 (SpaceXAI) 1577 · 2. Grok 4.5 (SpaceXAI) 1313 — full ranking of 2 models at https://aireleasetracker.com/benchmark/aa-briefcase - **GDPval (win/tie rate)**: 1. GPT-5.5 (OpenAI) 84.9% · 2. GPT-5.4 (OpenAI) 83% · 3. GPT-5.5-Pro (OpenAI) 82.3% · 4. GPT-5.4-Pro (OpenAI) 82% · 5. Claude Opus 4.7 (Anthropic) 80.3% — full ranking of 6 models at https://aireleasetracker.com/benchmark/gdpval-winrate - **CharXiv Reasoning**: 1. Muse Spark (Meta) 88.9% · 2. Muse Spark 1.1 (Meta) 88.4% · 2. Qwen3.8-Max (Qwen) 88.4% · 4. Kimi K3 (Moonshot AI) 84.8% · 5. Gemini 3.7 Flash (Google) 84.5% — full ranking of 14 models at https://aireleasetracker.com/benchmark/charxiv-reasoning - **GDP.PDF**: 1. Gemini 3.7 Flash (Google) 34% — full ranking of 1 models at https://aireleasetracker.com/benchmark/gdp-pdf - **LVBench**: 1. Gemini 3.7 Flash (Google) 85.4% — full ranking of 1 models at https://aireleasetracker.com/benchmark/lvbench - **BabyVision**: 1. Qwen3.8-Max (Qwen) 82% · 2. Muse Spark 1.1 (Meta) 76.3% · 3. Muse Spark (Meta) 39.9% — full ranking of 3 models at https://aireleasetracker.com/benchmark/babyvision - **MMMU-Pro**: 1. Gemini 3.5 Flash (Google) 83.6% · 2. Kimi K3 (Moonshot AI) 81.6% · 3. Gemini 3.0 Flash (Google) 81.2% · 3. GPT-5.5 (OpenAI) 81.2% · 5. Gemini 3.1 Pro (Google) 80.5% — full ranking of 10 models at https://aireleasetracker.com/benchmark/mmmu-pro - **MMMU**: 1. Qwen3.5 (Qwen) 85% · 2. Qwen3.6 (Qwen) 81.7% · 3. Gemini 3.0 Pro (Google) 81% · 4. Muse Spark (Meta) 80.4% · 5. GPT-5.1 (OpenAI) 76% — full ranking of 7 models at https://aireleasetracker.com/benchmark/mmmu - **Blueprint-Bench 2**: 1. GPT-5.5 (OpenAI) 36.2% · 2. Gemini 3.5 Flash (Google) 33.6% · 3. Gemini 3.1 Pro (Google) 26.5% · 4. Claude Opus 4.7 (Anthropic) 24.5% · 5. Claude Sonnet 4.6 (Anthropic) 6.7% — full ranking of 6 models at https://aireleasetracker.com/benchmark/blueprint-bench-2 - **MRCR v2 (8-needle) (128k average)**: 1. Gemini 3.7 Flash (Google) 97% · 2. GPT-5.5 (OpenAI) 94.8% · 3. Claude Sonnet 4.6 (Anthropic) 84.9% · 3. Gemini 3.1 Pro (Google) 84.9% · 5. Gemini 3.5 Flash (Google) 77.3% — full ranking of 7 models at https://aireleasetracker.com/benchmark/mrcr-v2-128k - **MRCR v2 (8-needle) (1M pointwise)**: 1. Gemini 3.7 Flash (Google) 62.5% · 2. Gemini 3.5 Flash (Google) 26.6% · 3. Gemini 3.1 Pro (Google) 26.3% · 4. Gemini 3.0 Flash (Google) 22.1% — full ranking of 4 models at https://aireleasetracker.com/benchmark/mrcr-v2-1m - **Arena Elo (Text)**: 1. Claude Fable 5 (Anthropic) 1509 · 2. Claude Opus 4.6 (Anthropic) 1504 · 3. Claude Opus 4.7 (Anthropic) 1503 · 4. Muse Spark 1.2 (Meta) 1498 · 5. Qwen3.8-Max (Qwen) 1497 — full ranking of 31 models at https://aireleasetracker.com/benchmark/arena-text-elo - **Arena Elo (Code)**: 1. Kimi K3 (Moonshot AI) 1679 · 2. Claude Opus 5 (Anthropic) 1673 · 3. Qwen3.8-Max (Qwen) 1667 · 4. Claude Fable 5 (Anthropic) 1631 · 5. GPT-5.6 Sol (OpenAI) 1620 — full ranking of 48 models at https://aireleasetracker.com/benchmark/arena-code-elo ## Labs tracked ### OpenAI Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. - Tracked releases: 43 - Official site: https://openai.com - Also known at: https://en.wikipedia.org/wiki/OpenAI, https://x.com/OpenAI, https://github.com/openai - Release history: https://aireleasetracker.com/company/openai ### Anthropic AI safety company building the Claude family of models. Founded in 2021 by former OpenAI researchers. - Tracked releases: 24 - Official site: https://www.anthropic.com - Also known at: https://en.wikipedia.org/wiki/Anthropic, https://x.com/AnthropicAI, https://github.com/anthropics - Release history: https://aireleasetracker.com/company/anthropic ### Google Builds the Gemini family of models through Google DeepMind. Integrates AI across Google products. - Tracked releases: 28 - Official site: https://deepmind.google - Also known at: https://en.wikipedia.org/wiki/Google_DeepMind, https://x.com/GoogleDeepMind - Release history: https://aireleasetracker.com/company/google ### Meta Develops the open-weight Llama series of models. Committed to open-source AI research. - Tracked releases: 14 - Official site: https://ai.meta.com - Also known at: https://en.wikipedia.org/wiki/Meta_AI, https://x.com/AIatMeta, https://github.com/meta-llama - Release history: https://aireleasetracker.com/company/meta ### SpaceXAI Elon Musk's AI company building the Grok series of models. Founded in 2023 as xAI, now part of SpaceX. - Tracked releases: 16 - Official site: https://x.ai - Also known at: https://en.wikipedia.org/wiki/XAI_(company), https://x.com/xai, https://github.com/xai-org - Release history: https://aireleasetracker.com/company/xai ### DeepSeek Chinese AI lab known for efficient, open-weight models. Gained attention for strong performance at lower cost. - Tracked releases: 22 - Official site: https://www.deepseek.com - Also known at: https://en.wikipedia.org/wiki/DeepSeek, https://x.com/deepseek_ai, https://github.com/deepseek-ai - Release history: https://aireleasetracker.com/company/deepseek ### Mistral French AI company building open and commercial models. Founded in 2023 by former Meta and DeepMind researchers. - Tracked releases: 32 - Official site: https://mistral.ai - Also known at: https://en.wikipedia.org/wiki/Mistral_AI, https://x.com/MistralAI, https://github.com/mistralai - Release history: https://aireleasetracker.com/company/mistral ### Moonshot AI Chinese AI company building the Kimi family of models. Founded in 2023, known for long-context and efficient MoE architectures. - Tracked releases: 12 - Official site: https://www.moonshot.ai - Also known at: https://en.wikipedia.org/wiki/Moonshot_AI, https://x.com/Kimi_Moonshot, https://github.com/MoonshotAI - Release history: https://aireleasetracker.com/company/moonshot ### Cursor - Tracked releases: 4 - Official site: https://cursor.com - Also known at: https://en.wikipedia.org/wiki/Cursor_(code_editor), https://x.com/cursor_ai - Release history: https://aireleasetracker.com/company/cursor ### Z.ai Chinese AI lab spun out of Tsinghua University (formerly Zhipu AI), building the open-weight GLM family. Rebranded internationally as Z.ai in 2025. - Tracked releases: 12 - Official site: https://z.ai - Also known at: https://en.wikipedia.org/wiki/Zhipu_AI, https://x.com/Zai_org, https://github.com/zai-org - Release history: https://aireleasetracker.com/company/zai ### Qwen Alibaba's AI lab, building the Qwen family. The most prolific publisher of open-weight models of any major lab, alongside a proprietary Max and Plus tier sold through Alibaba Cloud. - Tracked releases: 22 - Official site: https://qwen.ai - Also known at: https://en.wikipedia.org/wiki/Qwen, https://x.com/Alibaba_Qwen, https://github.com/QwenLM, https://huggingface.co/Qwen - Release history: https://aireleasetracker.com/company/qwen ## Every tracked release (newest first) ### GLM-5.3 - URL: https://aireleasetracker.com/model/zai/glm-5.3 - Developer: Z.ai - Release date: Friday, Aug 14 2026 - Licensing: Proprietary - Parameters: 743B - Benchmarks published at release: - DeepSWE 1.1: 66.9% - Terminal-Bench 3.0: 28.3% - CyberGym: 84.5% - ExploitBench: 54.4% - ExploitGym (6-hour budget): 130 - ExploitGym (2-hour budget): 105 - Humanity's Last Exam (with tools): 62.5% - Agent's Last Exam (pass@1): 28.5% - AutomationBench: 48.2% - GDPval-AA v2: 1769 GLM-5.3, released August 14, 2026, was Z.ai's coding-and-cybersecurity push — "Built to Code. Ready for Cyber Defense." — built by post-training the 743B GLM-5-generation base rather than scaling it up. At release it scored 66.9% on DeepSWE, more than twenty points above GLM-5.2, with 62.5% on Humanity's Last Exam with tools and a GDPval-AA v2 rating of 1769 — at the time the best published figure from any open-model lab, and ahead of the Claude and GPT flagships in Z.ai's own launch comparison. The cybersecurity positioning was the real novelty: 84.5% on CyberGym topped Z.ai's launch chart outright, and 54.4% on ExploitBench — CMU's exploit-development ladder — set a launch-time high among open-model labs, though the closed Anthropic and OpenAI flagships still led that test by over twenty points. Unusually for the GLM line, launch day came without weights: access started through the GLM Coding Plan and ZCode, with API access and open weights slated to follow in stages after safety evaluations. It was Z.ai's seventh flagship release in thirteen months. **Q: When was GLM-5.3 released?** A: GLM-5.3 was released by Z.ai on Friday, Aug 14 2026. **Q: Who made GLM-5.3?** A: GLM-5.3 was built by Z.ai. Chinese AI lab spun out of Tsinghua University (formerly Zhipu AI), building the open-weight GLM family. Rebranded internationally as Z.ai in 2025. **Q: What benchmark scores did GLM-5.3 get?** A: GLM-5.3 reports 10 tracked benchmark scores — DeepSWE 1.1: 66.9%; Terminal-Bench 3.0: 28.3%; CyberGym: 84.5%; ExploitBench: 54.4%; ExploitGym (6-hour budget): 130; ExploitGym (2-hour budget): 105; Humanity's Last Exam (with tools): 62.5%; Agent's Last Exam (pass@1): 28.5%; AutomationBench: 48.2%; GDPval-AA v2: 1769. Scores are the figures published at release by Z.ai. It holds the best score among all models tracked here on Terminal-Bench 3.0, ExploitBench, ExploitGym (6-hour budget), ExploitGym (2-hour budget), Agent's Last Exam (pass@1) and AutomationBench. **Q: How many parameters does GLM-5.3 have?** A: GLM-5.3 is reported at 743B parameters. **Q: Is GLM-5.3 open source?** A: No. GLM-5.3 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after GLM-5.3?** A: Z.ai's previous tracked release was GLM-5.2 on Jun 16 2026, 59 days earlier. It is the most recent Z.ai model tracked on AI Release Tracker. ### Qwen3.8-27B - URL: https://aireleasetracker.com/model/qwen/qwen3.8-27b - Developer: Qwen - Release date: Friday, Aug 14 2026 - Licensing: Open Weight - Parameters: 27B - Context window: 262k - Benchmarks published at release: - SWE-Bench Pro: 61.7% - DeepSWE 1.1: 42.2% - NL2Repo-Bench: 42.3% - QwenSWEBench: 79% - LiveCodeBench: 90.3% - Terminal-Bench 2.1: 73% - JobBench: 33.4% - CoWorkBench: 70.7% - Humanity's Last Exam (no tools): 30.8% - GPQA Diamond: 89.2% - IFBench: 79.5% - Agent's Last Exam (pass@1): 20.4% - Agent's Last Exam (score): 42.9% Qwen3.8-27B, released August 14, 2026, brought Alibaba's Qwen3.8 generation to a size intended for local deployment: a 27B-parameter dense, natively multimodal model with downloadable weights under Apache 2.0. Its 262K-token native context window could be extended to 1M through YaRN, giving builders much of the long-context reach of Qwen3.8-Max without the flagship's 2.4-trillion-parameter footprint or hosted-only multimodal product. The launch table concentrated on coding and office agents. Qwen3.8-27B scored 61.7% on SWE-Bench Pro, 42.2% on DeepSWE 1.1 and 79.0% on QwenSWEBench at release, all sizable gains over the earlier Qwen3.6-27B. Its 70.7% on CoWorkBench also edged the 68.2% Alibaba reported for Opus 4.6 Max, while 90.3% on LiveCodeBench v6 and 89.2% on GPQA Diamond put the compact model close to the larger proprietary systems in Qwen's comparison. **Q: When was Qwen3.8-27B released?** A: Qwen3.8-27B was released by Qwen on Friday, Aug 14 2026. **Q: Who made Qwen3.8-27B?** A: Qwen3.8-27B was built by Qwen. Alibaba's AI lab, building the Qwen family. The most prolific publisher of open-weight models of any major lab, alongside a proprietary Max and Plus tier sold through Alibaba Cloud. **Q: What benchmark scores did Qwen3.8-27B get?** A: Qwen3.8-27B reports 13 tracked benchmark scores — SWE-Bench Pro: 61.7%; DeepSWE 1.1: 42.2%; NL2Repo-Bench: 42.3%; QwenSWEBench: 79%; LiveCodeBench: 90.3%; Terminal-Bench 2.1: 73%; JobBench: 33.4%; CoWorkBench: 70.7%; Humanity's Last Exam (no tools): 30.8%; GPQA Diamond: 89.2%; IFBench: 79.5%; Agent's Last Exam (pass@1): 20.4%; Agent's Last Exam (score): 42.9%. Scores are the figures published at release by Qwen. It holds the best score among all models tracked here on NL2Repo-Bench, QwenSWEBench, CoWorkBench, IFBench and Agent's Last Exam (score). **Q: What is the context window of Qwen3.8-27B?** A: Qwen3.8-27B has a context window of 262k. That is the maximum amount of input plus output the model can hold in a single request. **Q: How many parameters does Qwen3.8-27B have?** A: Qwen3.8-27B is reported at 27B parameters. **Q: Is Qwen3.8-27B open source?** A: Partly. Qwen3.8-27B is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Qwen3.8-27B?** A: Qwen's previous tracked release was Qwen3.8-Max on Aug 3 2026, 11 days earlier. It is the most recent Qwen model tracked on AI Release Tracker. ### Gemini 3.7 Flash - URL: https://aireleasetracker.com/model/google/gemini-3.7-flash - Developer: Google - Release date: Thursday, Aug 13 2026 - Licensing: Proprietary - Context window: 1M - Benchmarks published at release: - DeepSWE 1.1: 65.3% - FrontierCode v1.1 (Main) (main split): 43.6% - Terminal-Bench 3.0: 14.9% - Terminal-Bench 2.1: 85.8% - Humanity's Last Exam (Verified): 53.6% - BioMysteryBench (hard): 43.5% - BioMysteryBench (human solved): 87.1% - LAB-Bench 2: 82.1% - OSWorld 2.0: 38.1% - Agent's Last Exam (pass@1): 26.3% - AutomationBench: 30.4% - Harvey's Legal Agent Benchmark: 90.7% - AA Intelligence Index: 56 - GDPval-AA v2: 1525 - CharXiv Reasoning: 84.5% - GDP.PDF: 34% - LVBench: 85.4% - MRCR v2 (8-needle) (128k average): 97% - MRCR v2 (8-needle) (1M pointwise): 62.5% - Arena Elo (Code): 1588 Gemini 3.7 Flash, released August 13, 2026, arrived three weeks after Gemini 3.6 Flash and moved the biggest numbers of the whole Flash line in coding. At launch it scored 65.3% on DeepSWE v1.1 for long-horizon software engineering, against 49.0% for 3.6 Flash, and 43.6% on the main split of FrontierCode 1.1 — the best figure in Google's launch comparison, ahead of Claude Sonnet 5 and GPT-5.6 Terra. Web development moved with it, to an Arena.ai WebDev Elo of 1588, and the gains extended past code: 30.4% on AutomationBench for enterprise workflows (up from 17.0%), 34.0% on GDP.PDF document comprehension (up from 22.0%), 90.7% on Harvey's LAB-AA legal benchmark, and 97.0% on MRCR v2 128k long-context recall. On the Artificial Analysis Intelligence Index it landed at 56, above Sonnet 5 at 55 and a point below GPT-5.6 Terra and Muse Spark 1.2. Price was the other half of the pitch. Google launched it at an introductory $0.75 per million input tokens and $3.75 per million output through the end of 2026 — half what 3.6 Flash had cost at its own launch — with the rate reverting to $1.50/$7.50 in January 2027. It shipped with a 1M-token context window and multimodal input covering images, video, audio, and PDFs, across the Gemini API, AI Studio, Antigravity, Android Studio, and Gemini Enterprise, and reached consumers through Spark for AI Pro and Ultra subscribers. The agentic computer-use scores were the exception to the sweep: 38.1% on OSWorld 2.0 and 14.9% on Terminal-Bench 3.0 both trailed GPT-5.6 Terra at release. Three Flash-tier upgrades between May and August, against a Pro flagship untouched since February, made clear which tier Google saw as the centre of its lineup. **Q: When was Gemini 3.7 Flash released?** A: Gemini 3.7 Flash was released by Google on Thursday, Aug 13 2026. **Q: Who made Gemini 3.7 Flash?** A: Gemini 3.7 Flash was built by Google. Builds the Gemini family of models through Google DeepMind. Integrates AI across Google products. **Q: What benchmark scores did Gemini 3.7 Flash get?** A: Gemini 3.7 Flash reports 20 tracked benchmark scores — DeepSWE 1.1: 65.3%; FrontierCode v1.1 (Main) (main split): 43.6%; Terminal-Bench 3.0: 14.9%; Terminal-Bench 2.1: 85.8%; Humanity's Last Exam (Verified): 53.6%; BioMysteryBench (hard): 43.5%; BioMysteryBench (human solved): 87.1%; LAB-Bench 2: 82.1%; OSWorld 2.0: 38.1%; Agent's Last Exam (pass@1): 26.3%; AutomationBench: 30.4%; Harvey's Legal Agent Benchmark: 90.7%; AA Intelligence Index: 56; GDPval-AA v2: 1525; CharXiv Reasoning: 84.5%; GDP.PDF: 34%; LVBench: 85.4%; MRCR v2 (8-needle) (128k average): 97%; MRCR v2 (8-needle) (1M pointwise): 62.5%; Arena Elo (Code): 1588. Scores are the figures published at release by Google. It holds the best score among all models tracked here on Humanity's Last Exam (Verified), LAB-Bench 2, Harvey's Legal Agent Benchmark, GDP.PDF, LVBench, MRCR v2 (8-needle) (128k average) and MRCR v2 (8-needle) (1M pointwise). **Q: What is the context window of Gemini 3.7 Flash?** A: Gemini 3.7 Flash has a context window of 1M. That is the maximum amount of input plus output the model can hold in a single request. **Q: Is Gemini 3.7 Flash open source?** A: No. Gemini 3.7 Flash is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Gemini 3.7 Flash?** A: Google's previous tracked release was Gemini 3.5 Flash Cyber on Jul 21 2026, 23 days earlier. It is the most recent Google model tracked on AI Release Tracker. ### DeepSeek-V4-Pro-0813 - URL: https://aireleasetracker.com/model/deepseek/deepseek-v4-pro-0813 - Developer: DeepSeek - Release date: Thursday, Aug 13 2026 - Licensing: Proprietary - Benchmarks published at release: - Terminal-Bench 2.1: 87.9% - Toolathlon-Verified: 74.1% - CyberGym: 83.3% - Humanity's Last Exam (no tools): 42.7% - Humanity's Last Exam (with tools): 60% - AutomationBench: 31.8% DeepSeek-V4-Pro-0813, released August 13, 2026, brought the top tier of the V4 line up to the agentic post-training that had landed on Flash two weeks earlier. On DeepSeek's own harness it scored 87.9 on Terminal-Bench 2.1, against 82.7 for V4-Flash-0731 and 72.1 for the April V4-Pro-Preview build, and 83.3 on CyberGym — the highest figure in the lab's launch comparison at the time, marginally ahead of Claude Fable 5. It also posted 42.7% on Humanity's Last Exam without tools and 60.0% with them, 74.1 on Toolathlon-Verified and 31.8 on the public AutomationBench split. As with the July Flash release, DeepSeek ran the code-agent evaluations through its own unreleased harness in minimal mode, and its numbers for rival models did not always match those labs' published figures, so the chart reads best as a within-family comparison. The headline changes were operational rather than architectural. V4-Pro and V4-Flash both gained a selectable reasoning effort — low for simple prompts, high for everyday agent work, max for hard tasks — replacing the single fixed thinking budget of the earlier builds, and the release added native OpenAI Responses API support with one-click Codex setup. Distribution followed the pattern the Flash refresh established: the checkpoint reached the app and web product under an "Expert Mode" toggle and shipped through the API under unchanged model names, so existing callers were moved onto it without a code change. DeepSeek's launch post listed app, web, and API availability only, and announced no accompanying weight release. **Q: When was DeepSeek-V4-Pro-0813 released?** A: DeepSeek-V4-Pro-0813 was released by DeepSeek on Thursday, Aug 13 2026. **Q: Who made DeepSeek-V4-Pro-0813?** A: DeepSeek-V4-Pro-0813 was built by DeepSeek. Chinese AI lab known for efficient, open-weight models. Gained attention for strong performance at lower cost. **Q: What benchmark scores did DeepSeek-V4-Pro-0813 get?** A: DeepSeek-V4-Pro-0813 reports 6 tracked benchmark scores — Terminal-Bench 2.1: 87.9%; Toolathlon-Verified: 74.1%; CyberGym: 83.3%; Humanity's Last Exam (no tools): 42.7%; Humanity's Last Exam (with tools): 60%; AutomationBench: 31.8%. Scores are the figures published at release by DeepSeek. **Q: Is DeepSeek-V4-Pro-0813 open source?** A: No. DeepSeek-V4-Pro-0813 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after DeepSeek-V4-Pro-0813?** A: DeepSeek's previous tracked release was DeepSeek-V4-Flash-0731 on Jul 31 2026, 13 days earlier. It is the most recent DeepSeek model tracked on AI Release Tracker. ### Grok 4.6 - URL: https://aireleasetracker.com/model/xai/grok-4.6 - Developer: SpaceXAI - Release date: Wednesday, Aug 12 2026 - Licensing: Proprietary - Benchmarks published at release: - CursorBench v3.2: 69.9% - DeepSWE 1.1: 65.9% - FrontierCode v1.1 (Extended) (extended split): 61.3% - APEX-SWE: 56.4% - Next.js Evals: 92% - Terminal-Bench 3.0: 26% - APEX-Agents: 57.5% - Harvey's Legal Agent Benchmark: 15.8% - AA Intelligence Index: 61 - GDPval-AA v2: 1753 - AA-Briefcase: 1577 Grok 4.6, released August 12, 2026, was xAI's bet that post-training alone could deliver a flagship upgrade. It reused the 1.5-trillion-parameter base model of Grok 4.5, five weeks its senior, and poured the gains into supervised fine-tuning on regenerated trajectories and wide-ranging reinforcement learning across engineering and domain-specific environments. At release it scored 61 on the Artificial Analysis Intelligence Index — up from 56 for Grok 4.5, and level with GPT-5.6 Sol at the time — while priced at $2 per million input tokens and $6 per million output. xAI pitched it at long-running agents and knowledge work rather than raw coding leaderboards. At launch it posted a GDPval-AA v2 Elo of 1753, 15.8% on Harvey's Legal Agent Benchmark — the top score in xAI's own launch comparison — and 69.9% on CursorBench v3.2, within a point of Claude Fable 5's published number at the time. Terminal work stayed a relative weakness, at 26% on the much harder Terminal-Bench 3.0. Elon Musk said the larger Grok 4.7 would follow within weeks, trading some serving speed for capability. **Q: When was Grok 4.6 released?** A: Grok 4.6 was released by SpaceXAI on Wednesday, Aug 12 2026. **Q: Who made Grok 4.6?** A: Grok 4.6 was built by SpaceXAI. Elon Musk's AI company building the Grok series of models. Founded in 2023 as xAI, now part of SpaceX. **Q: What benchmark scores did Grok 4.6 get?** A: Grok 4.6 reports 11 tracked benchmark scores — CursorBench v3.2: 69.9%; DeepSWE 1.1: 65.9%; FrontierCode v1.1 (Extended) (extended split): 61.3%; APEX-SWE: 56.4%; Next.js Evals: 92%; Terminal-Bench 3.0: 26%; APEX-Agents: 57.5%; Harvey's Legal Agent Benchmark: 15.8%; AA Intelligence Index: 61; GDPval-AA v2: 1753; AA-Briefcase: 1577. Scores are the figures published at release by SpaceXAI. It holds the best score among all models tracked here on FrontierCode v1.1 (Extended) (extended split), APEX-SWE, APEX-Agents, AA Intelligence Index and AA-Briefcase. **Q: Is Grok 4.6 open source?** A: No. Grok 4.6 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Grok 4.6?** A: SpaceXAI's previous tracked release was Grok 4.5 on Jul 8 2026, 35 days earlier. It is the most recent SpaceXAI model tracked on AI Release Tracker. ### Muse Glimmer - URL: https://aireleasetracker.com/model/meta/muse-glimmer - Developer: Meta - Release date: Monday, Aug 10 2026 - Licensing: Open Weight - Parameters: 30B - Benchmarks published at release: - SWE-Bench Pro: 51.2% - SWE-Bench Verified: 76% - Terminal-Bench 2.1: 51.7% - MCP Atlas: 75.5% - Humanity's Last Exam (no tools): 22% - GPQA Diamond: 83.5% - OSWorld-Verified: 65.9% - GDPval-AA v2: 953 - CharXiv Reasoning: 78.8% - MMMU-Pro: 74% Muse Glimmer, released August 10, 2026, was the first open model from Meta Superintelligence Labs and the end of the closed stretch that began with Muse Spark in April: a 30B dense model distilled from Muse Spark 1.2, published on Hugging Face under Apache 2.0. Meta pitched it at always-on local agents rather than raw capability — small enough to run on a single consumer GPU or a Mac, multimodal through a dedicated perception encoder, trained on more than 100 languages, and tuned for the parts of agent work that break in practice: tool calls with exact schemas, plans held over long horizons, and recovery after a tool returns something unexpected. The engineering around it aimed at the same target, with 4-bit quantization bringing memory under 20GB and DFlash speculative decoding cutting generation time by 3.1x on an RTX 5090 and 1.8x on an M5 Max. Meta benchmarked it against the open models in its size class, Gemma 4 31B and Qwen3.6 27B, and led on agentic tool use: 75.5% on MCP Atlas at release against 54.2% and 62.5% for the other two, plus 74.6% on DeepSearch QA and 47.6% on WildClawBench. It scored 76.0% on SWE-Bench Verified, 51.2% on SWE-Bench Pro, 83.5% on GPQA Diamond, and 94.7% on AIME 2026, while Qwen3.6 took the computer-use and long-form coding rows — 75.6% to Muse Glimmer's 65.9% on OSWorld-Verified, 60.7% to 51.7% on Terminal-Bench 2.1. Unlike the Muse Spark 1.2 launch, Meta published a methodology report naming every harness and judge, though most of the comparison numbers were still its own reproductions rather than each model's self-reported best. Meta said open weights for a version of Muse Spark 1.2 would follow. **Q: When was Muse Glimmer released?** A: Muse Glimmer was released by Meta on Monday, Aug 10 2026. **Q: Who made Muse Glimmer?** A: Muse Glimmer was built by Meta. Develops the open-weight Llama series of models. Committed to open-source AI research. **Q: What benchmark scores did Muse Glimmer get?** A: Muse Glimmer reports 10 tracked benchmark scores — SWE-Bench Pro: 51.2%; SWE-Bench Verified: 76%; Terminal-Bench 2.1: 51.7%; MCP Atlas: 75.5%; Humanity's Last Exam (no tools): 22%; GPQA Diamond: 83.5%; OSWorld-Verified: 65.9%; GDPval-AA v2: 953; CharXiv Reasoning: 78.8%; MMMU-Pro: 74%. Scores are the figures published at release by Meta. **Q: How many parameters does Muse Glimmer have?** A: Muse Glimmer is reported at 30B parameters. **Q: Is Muse Glimmer open source?** A: Partly. Muse Glimmer is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Muse Glimmer?** A: Meta's previous tracked release was Muse Spark 1.2 on Aug 5 2026, 5 days earlier. It is the most recent Meta model tracked on AI Release Tracker. ### GPT-5.6-Cyber - URL: https://aireleasetracker.com/model/openai/gpt-5.6-cyber - Developer: OpenAI - Release date: Monday, Aug 10 2026 - Licensing: Proprietary - Benchmarks: none published in this dataset GPT-5.6-Cyber, released August 10, 2026, was OpenAI's most capable and most permissive security model at the time: a GPT-5.6 Sol derivative trained for zero-day discovery and exploit-chain development, and tuned to answer the sensitive queries general models refuse. At launch OpenAI reported it completed 95% of advanced cybersecurity requests, against 1.5% for GPT-5.6 Sol and 57.3% for its predecessor GPT-5.5-Cyber; in pre-release testing it found two previously unknown, chainable vulnerabilities in Chrome's V8 engine. OpenAI rated the model High for cybersecurity under its Preparedness Framework, below the Critical threshold, and restructured Daybreak around it: a Blue tier offering GPT-5.6 Sol without system-level cyber guardrails for defensive work, and a Red tier gating GPT-5.6-Cyber itself behind identity verification, monitoring, and legal declarations for vulnerability research and exploit validation. Partners including CrowdStrike, Cisco, IBM, and Palo Alto Networks were cleared to build the models into security products at launch. **Q: When was GPT-5.6-Cyber released?** A: GPT-5.6-Cyber was released by OpenAI on Monday, Aug 10 2026. **Q: Who made GPT-5.6-Cyber?** A: GPT-5.6-Cyber was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: Is GPT-5.6-Cyber open source?** A: No. GPT-5.6-Cyber is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after GPT-5.6-Cyber?** A: OpenAI's previous tracked release was GPT-5.6 Luna on Jun 26 2026, 45 days earlier. It is the most recent OpenAI model tracked on AI Release Tracker. ### Muse Spark 1.2 - URL: https://aireleasetracker.com/model/meta/muse-spark-1.2 - Developer: Meta - Release date: Wednesday, Aug 5 2026 - Licensing: Proprietary - Context window: 1M - Benchmarks published at release: - DeepSWE 1.1: 59.3% - Terminal-Bench 2.1: 82.9% - Arena Elo (Text): 1498 - Arena Elo (Code): 1544 Muse Spark 1.2, released August 5, 2026, was a coding-focused update to Muse Spark 1.1 that Meta shipped alongside Muse Code, its first coding agent — a terminal tool for macOS and Linux built around persistent background agents and parallel sub-agents in isolated git worktrees. Meta said the two were co-trained, which framed the release as an agent-plus-model product rather than a general model refresh: the pitch was repository-scale software engineering, with the model tuned for the harness it ships in. The launch numbers were narrow and squarely on coding. Muse Spark 1.2 scored 82.9% on Terminal-Bench 2.1 and 59.3% on DeepSWE 1.1, in both cases the second-best figure on Meta's own comparison table behind Claude Opus 5 — a considerably tighter gap to the frontier than Muse Spark had managed at its April debut, when it posted 10.0% on DeepSWE 1.1. It carried a 1M-token context window and shipped proprietary, served through the Meta Model API at $1.25 per million input tokens against $4.25 per million output, roughly a third of what the flagship models it was benchmarked against charged at the time. Meta published the results as charts without a methodology write-up, so the comparison figures for rival models came from its own harness rather than each lab's. Five days later Meta distilled it into Muse Glimmer and put the result on Hugging Face, its first open release since Llama 4. **Q: When was Muse Spark 1.2 released?** A: Muse Spark 1.2 was released by Meta on Wednesday, Aug 5 2026. **Q: Who made Muse Spark 1.2?** A: Muse Spark 1.2 was built by Meta. Develops the open-weight Llama series of models. Committed to open-source AI research. **Q: What benchmark scores did Muse Spark 1.2 get?** A: Muse Spark 1.2 reports 4 tracked benchmark scores — DeepSWE 1.1: 59.3%; Terminal-Bench 2.1: 82.9%; Arena Elo (Text): 1498; Arena Elo (Code): 1544. Scores are the figures published at release by Meta. **Q: What is the context window of Muse Spark 1.2?** A: Muse Spark 1.2 has a context window of 1M. That is the maximum amount of input plus output the model can hold in a single request. **Q: Is Muse Spark 1.2 open source?** A: No. Muse Spark 1.2 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Muse Spark 1.2?** A: Meta's previous tracked release was Muse Spark 1.1 on Jul 9 2026, 27 days earlier. It was followed by Muse Glimmer on Aug 10 2026. ### Qwen3.8-Max - URL: https://aireleasetracker.com/model/qwen/qwen3.8-max - Developer: Qwen - Release date: Monday, Aug 3 2026 - Licensing: Proprietary - Parameters: 2.4T - Context window: 1M - Benchmarks published at release: - SWE-Bench Pro: 67.7% - PaperBench: 93% - Terminal-Bench 2.1: 86.6% - JobBench: 53.4% - OSWorld-Verified: 86.1% - CharXiv Reasoning: 88.4% - BabyVision: 82% - Arena Elo (Text): 1497 - Arena Elo (Code): 1667 Qwen3.8-Max, released August 3, 2026, was the largest model Alibaba had built: 2.4 trillion total parameters with 95B active, a 1M-token context window, and hybrid thinking. Alibaba previewed it on July 19 and shipped it to the Alibaba Cloud API two weeks later, placing it fifth on Text Arena and second on Vision Arena at launch. The launch numbers pointed at agents rather than raw coding. It scored 93.0% on PaperBench and 86.1% on OSWorld-Verified — both ahead of every model Alibaba benchmarked against, including GPT-5.6 Sol and Claude Fable 5 — along with 86.6% on Terminal-Bench 2.1 and 53.4% on JobBench. On conventional software engineering it was further back, at 67.7% on SWE-Bench Pro against the 80.3% Claude Fable 5 had posted at its own release two months earlier. Multimodal results were strong without a code tool: 88.4% on CharXiv chart reasoning and 82.0% on BabyVision. Alibaba announced at release that it intended to publish the weights, and on August 12, 2026 it opened the text-only base model underneath the product, Qwen3.8-2.4T-A95B, under a custom licence — by a wide margin the largest weights any lab had published to that point, though the hosted Max kept vision input, non-thinking mode and the full 1M context for itself. Opening a Max-tier flagship cut against the pattern the Max line had followed since Qwen3-Max, and the smaller Apache-licensed Qwen3.8-27B followed two days later. The launch landed two and a half weeks after Moonshot AI's Kimi K3, in a stretch where the Chinese labs were shipping trillion-parameter models within weeks of each other. **Q: When was Qwen3.8-Max released?** A: Qwen3.8-Max was released by Qwen on Monday, Aug 3 2026. **Q: Who made Qwen3.8-Max?** A: Qwen3.8-Max was built by Qwen. Alibaba's AI lab, building the Qwen family. The most prolific publisher of open-weight models of any major lab, alongside a proprietary Max and Plus tier sold through Alibaba Cloud. **Q: What benchmark scores did Qwen3.8-Max get?** A: Qwen3.8-Max reports 9 tracked benchmark scores — SWE-Bench Pro: 67.7%; PaperBench: 93%; Terminal-Bench 2.1: 86.6%; JobBench: 53.4%; OSWorld-Verified: 86.1%; CharXiv Reasoning: 88.4%; BabyVision: 82%; Arena Elo (Text): 1497; Arena Elo (Code): 1667. Scores are the figures published at release by Qwen. It holds the best score among all models tracked here on PaperBench, OSWorld-Verified and BabyVision. **Q: What is the context window of Qwen3.8-Max?** A: Qwen3.8-Max has a context window of 1M. That is the maximum amount of input plus output the model can hold in a single request. **Q: How many parameters does Qwen3.8-Max have?** A: Qwen3.8-Max is reported at 2.4T parameters. **Q: Is Qwen3.8-Max open source?** A: No. Qwen3.8-Max is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Qwen3.8-Max?** A: Qwen's previous tracked release was Qwen3.7-Plus on Jun 1 2026, 63 days earlier. It was followed by Qwen3.8-27B on Aug 14 2026. ### DeepSeek-V4-Flash-0731 - URL: https://aireleasetracker.com/model/deepseek/deepseek-v4-flash-0731 - Developer: DeepSeek - Release date: Friday, Jul 31 2026 - Licensing: Proprietary - Benchmarks published at release: - BullshitBench v2: 38% - Terminal-Bench 2.1: 82.7% - Toolathlon-Verified: 70.3% - CyberGym: 76.7% - AutomationBench: 25.1% DeepSeek-V4-Flash-0731, announced July 31, 2026 as the "official" V4-Flash API in public beta, was a post-training refresh of the April V4-Flash build — same architecture, retrained for agentic work. The jump was unusually large for a checkpoint update: on DeepSeek's own evaluation harness it scored 82.7 on Terminal-Bench 2.1 against 61.8 for the April build, and 76.7 on CyberGym, surpassing the larger V4-Pro-Preview on several coding benchmarks despite being the cheaper tier. DeepSeek's harness numbers for competitor models differed from those labs' own published figures, so its launch chart is best read as within-family comparison. The release doubled as a renaming: DeepSeek retroactively designated the April 24 builds V4-Flash-Preview and V4-Pro-Preview, and served the new checkpoint under the unchanged deepseek-v4-flash API name — existing callers were switched to it without a code change. It added native support for the OpenAI Responses API format and Codex compatibility. Unlike every prior DeepSeek release, it shipped API-only at launch: no 0731 weights were published, and Hugging Face still carried the April build at release. V4-Pro, the app, and the web product stayed on their April checkpoints until the 0813 Pro refresh two weeks later. **Q: When was DeepSeek-V4-Flash-0731 released?** A: DeepSeek-V4-Flash-0731 was released by DeepSeek on Friday, Jul 31 2026. **Q: Who made DeepSeek-V4-Flash-0731?** A: DeepSeek-V4-Flash-0731 was built by DeepSeek. Chinese AI lab known for efficient, open-weight models. Gained attention for strong performance at lower cost. **Q: What benchmark scores did DeepSeek-V4-Flash-0731 get?** A: DeepSeek-V4-Flash-0731 reports 5 tracked benchmark scores — BullshitBench v2: 38%; Terminal-Bench 2.1: 82.7%; Toolathlon-Verified: 70.3%; CyberGym: 76.7%; AutomationBench: 25.1%. Scores are the figures published at release by DeepSeek. **Q: Is DeepSeek-V4-Flash-0731 open source?** A: No. DeepSeek-V4-Flash-0731 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after DeepSeek-V4-Flash-0731?** A: DeepSeek's previous tracked release was DeepSeek-V4-Flash on Apr 24 2026, 98 days earlier. It was followed by DeepSeek-V4-Pro-0813 on Aug 13 2026. ### Claude Opus 5 - URL: https://aireleasetracker.com/model/anthropic/claude-opus-5 - Developer: Anthropic - Release date: Friday, Jul 24 2026 - Licensing: Proprietary - Context window: 1M - Benchmarks published at release: - BullshitBench v2: 73% - Gray Swan IPI (k = 1): 0.2% - Gray Swan IPI (k = 10): 1.6% - Gray Swan IPI (k = 15): 2% - CursorBench v3.2: 70% - DeepSWE 1.1: 68.8% - FrontierCode v1.1 (Main) (main split): 53.4% - Next.js Evals: 88% - Supabase Evals (with skills): 95.5% - Supabase Evals (no skills): 90.9% - Frontier-Bench v0.1: 43.3% - BrowseComp: 90.8% - Humanity's Last Exam (no tools): 56.3% - Humanity's Last Exam (with tools): 64.7% - ARC-AGI-3: 30.2% - BioMysteryBench (hard): 49.4% - BioMysteryBench (human solved): 90.1% - OSWorld 2.0: 70.6% - AutomationBench: 26% - Harvey's Legal Agent Benchmark (Held-out): 11.7% - HealthBench Professional: 59.8% - GDPval-AA v2: 1861 - Arena Elo (Text): 1495 - Arena Elo (Code): 1673 Claude Opus 5, released July 24, 2026, brought Anthropic's Claude 5 generation to the Opus tier six weeks after Claude Fable 5 opened it — at half Fable's price, keeping Opus 4.8's $5 per million input tokens and $25 per million output. The pitch was flagship-class agentic capability at workhorse pricing: at launch it scored 43.3% on Frontier-Bench v0.1, more than double Opus 4.8's 21.1% and nearly ten points clear of Fable 5, and posted a GDPval-AA v2 Elo of 1861 for knowledge work, the best published score at the time. The launch card leaned on breadth: 90.8% on BrowseComp for agentic search, 70.6% on OSWorld 2.0 computer use, 64.7% on Humanity's Last Exam with tools, and 30.2% on ARC-AGI-3 — which Anthropic reported as roughly three times the next best published result on the novel problem-solving benchmark. Anthropic made it the default model on Claude Max and the strongest model available on Claude Pro, positioning Opus 5 as the everyday frontier model while Fable 5 kept the edge on a handful of evaluations, including DeepSWE agentic coding and Harvey's held-out legal benchmark. **Q: When was Claude Opus 5 released?** A: Claude Opus 5 was released by Anthropic on Friday, Jul 24 2026. **Q: Who made Claude Opus 5?** A: Claude Opus 5 was built by Anthropic. AI safety company building the Claude family of models. Founded in 2021 by former OpenAI researchers. **Q: What benchmark scores did Claude Opus 5 get?** A: Claude Opus 5 reports 24 tracked benchmark scores — BullshitBench v2: 73%; Gray Swan IPI (k = 1): 0.2%; Gray Swan IPI (k = 10): 1.6%; Gray Swan IPI (k = 15): 2%; CursorBench v3.2: 70%; DeepSWE 1.1: 68.8%; FrontierCode v1.1 (Main) (main split): 53.4%; Next.js Evals: 88%; Supabase Evals (with skills): 95.5%; Supabase Evals (no skills): 90.9%; Frontier-Bench v0.1: 43.3%; BrowseComp: 90.8%; Humanity's Last Exam (no tools): 56.3%; Humanity's Last Exam (with tools): 64.7%; ARC-AGI-3: 30.2%; BioMysteryBench (hard): 49.4%; BioMysteryBench (human solved): 90.1%; OSWorld 2.0: 70.6%; AutomationBench: 26%; Harvey's Legal Agent Benchmark (Held-out): 11.7%; HealthBench Professional: 59.8%; GDPval-AA v2: 1861; Arena Elo (Text): 1495; Arena Elo (Code): 1673. Scores are the figures published at release by Anthropic. It holds the best score among all models tracked here on Gray Swan IPI (k = 1), Gray Swan IPI (k = 10), Gray Swan IPI (k = 15), FrontierCode v1.1 (Main) (main split), Frontier-Bench v0.1, Humanity's Last Exam (no tools), Humanity's Last Exam (with tools), ARC-AGI-3, BioMysteryBench (hard), BioMysteryBench (human solved), OSWorld 2.0, Harvey's Legal Agent Benchmark (Held-out), HealthBench Professional and GDPval-AA v2. **Q: What is the context window of Claude Opus 5?** A: Claude Opus 5 has a context window of 1M. That is the maximum amount of input plus output the model can hold in a single request. **Q: Is Claude Opus 5 open source?** A: No. Claude Opus 5 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Claude Opus 5?** A: Anthropic's previous tracked release was Claude Sonnet 5 on Jun 30 2026, 24 days earlier. It is the most recent Anthropic model tracked on AI Release Tracker. ### Gemini 3.6 Flash - URL: https://aireleasetracker.com/model/google/gemini-3.6-flash - Developer: Google - Release date: Tuesday, Jul 21 2026 - Licensing: Proprietary - Benchmarks published at release: - BullshitBench v2: 39% - Gray Swan IPI (k = 1): 7.3% - Gray Swan IPI (k = 10): 32.2% - Gray Swan IPI (k = 15): 37.3% - CursorBench v3.2: 53.5% - DeepSWE 1.1: 49% - MLE-Bench: 63.9% - BU Bench: 68% - OSWorld-Verified: 83% - GDPval-AA v2: 1421 - Arena Elo (Text): 1482 - Arena Elo (Code): 1533 Gemini 3.6 Flash, released July 21, 2026 alongside Gemini 3.5 Flash-Lite and the security-focused Gemini 3.5 Flash Cyber, was pitched on efficiency rather than raw scale: Google priced it identically to Gemini 3.5 Flash ($1.50 per million input tokens, $7.50 per million output) while reporting it used roughly 17% fewer output tokens to deliver better results. At launch it posted 49% on DeepSWE 1.1 for long-horizon software engineering — up from 37% for 3.5 Flash — 63.9% on MLE-Bench for machine-learning engineering, a GDPval-AA v2 Elo of 1421 for knowledge work, and 83.0% on OSWorld-Verified computer use. The release continued Google's pattern of iterating fastest on its workhorse Flash tier, arriving just two months after Gemini 3.5 Flash and jumping the line's numbering to 3.6 while the Pro flagship stayed on 3.1. The pitch — a straight quality upgrade at the exact same cost — targeted the high-volume agentic workloads where Flash had become one of the most heavily used API models. Its turn at the top of the line was the shortest yet: Gemini 3.7 Flash replaced it three weeks later, on August 13, 2026. **Q: When was Gemini 3.6 Flash released?** A: Gemini 3.6 Flash was released by Google on Tuesday, Jul 21 2026. **Q: Who made Gemini 3.6 Flash?** A: Gemini 3.6 Flash was built by Google. Builds the Gemini family of models through Google DeepMind. Integrates AI across Google products. **Q: What benchmark scores did Gemini 3.6 Flash get?** A: Gemini 3.6 Flash reports 12 tracked benchmark scores — BullshitBench v2: 39%; Gray Swan IPI (k = 1): 7.3%; Gray Swan IPI (k = 10): 32.2%; Gray Swan IPI (k = 15): 37.3%; CursorBench v3.2: 53.5%; DeepSWE 1.1: 49%; MLE-Bench: 63.9%; BU Bench: 68%; OSWorld-Verified: 83%; GDPval-AA v2: 1421; Arena Elo (Text): 1482; Arena Elo (Code): 1533. Scores are the figures published at release by Google. It holds the best score among all models tracked here on MLE-Bench. **Q: Is Gemini 3.6 Flash open source?** A: No. Gemini 3.6 Flash is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Gemini 3.6 Flash?** A: Google's previous tracked release was Gemini 3.5 Flash on May 19 2026, 63 days earlier. It was followed by Gemini 3.5 Flash-Lite on Jul 21 2026. ### Gemini 3.5 Flash-Lite - URL: https://aireleasetracker.com/model/google/gemini-3.5-flash-lite - Developer: Google - Release date: Tuesday, Jul 21 2026 - Licensing: Proprietary - Benchmarks published at release: - BullshitBench v2: 65% - SWE-Bench Pro: 54.2% - Terminal-Bench 2.1: 54% - BU Bench: 49% - OSWorld-Verified: 74% - GDPval-AA v2: 1140 - Arena Elo (Text): 1459 - Arena Elo (Code): 1453 **Q: When was Gemini 3.5 Flash-Lite released?** A: Gemini 3.5 Flash-Lite was released by Google on Tuesday, Jul 21 2026. **Q: Who made Gemini 3.5 Flash-Lite?** A: Gemini 3.5 Flash-Lite was built by Google. Builds the Gemini family of models through Google DeepMind. Integrates AI across Google products. **Q: What benchmark scores did Gemini 3.5 Flash-Lite get?** A: Gemini 3.5 Flash-Lite reports 8 tracked benchmark scores — BullshitBench v2: 65%; SWE-Bench Pro: 54.2%; Terminal-Bench 2.1: 54%; BU Bench: 49%; OSWorld-Verified: 74%; GDPval-AA v2: 1140; Arena Elo (Text): 1459; Arena Elo (Code): 1453. Scores are the figures published at release by Google. **Q: Is Gemini 3.5 Flash-Lite open source?** A: No. Gemini 3.5 Flash-Lite is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Gemini 3.5 Flash-Lite?** A: Google's previous tracked release was Gemini 3.6 Flash on Jul 21 2026. It was followed by Gemini 3.5 Flash Cyber on Jul 21 2026. ### Gemini 3.5 Flash Cyber - URL: https://aireleasetracker.com/model/google/gemini-3.5-flash-cyber - Developer: Google - Release date: Tuesday, Jul 21 2026 - Licensing: Proprietary - Benchmarks: none published in this dataset Gemini 3.5 Flash Cyber, announced July 21, 2026 alongside Gemini 3.6 Flash and 3.5 Flash-Lite, was Google's first cybersecurity-specialised Gemini model — built to find and patch critical software vulnerabilities, and deployed inside Google DeepMind's CodeMender agent. Google reported competitive performance on CyberGym, the benchmark for discovering and exploiting software weaknesses in sandboxed environments, at launch. Unlike its general-purpose siblings, Flash Cyber did not ship to the public API: access was restricted to a limited pilot program for governments and trusted partners, reflecting the dual-use sensitivity of automated vulnerability discovery. The release marked the first time Google spun a domain-specialised variant out of the Flash line, positioning defensive security as its own product track rather than a capability folded into the flagship. **Q: When was Gemini 3.5 Flash Cyber released?** A: Gemini 3.5 Flash Cyber was released by Google on Tuesday, Jul 21 2026. **Q: Who made Gemini 3.5 Flash Cyber?** A: Gemini 3.5 Flash Cyber was built by Google. Builds the Gemini family of models through Google DeepMind. Integrates AI across Google products. **Q: Is Gemini 3.5 Flash Cyber open source?** A: No. Gemini 3.5 Flash Cyber is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Gemini 3.5 Flash Cyber?** A: Google's previous tracked release was Gemini 3.5 Flash-Lite on Jul 21 2026. It was followed by Gemini 3.7 Flash on Aug 13 2026. ### Kimi K3 - URL: https://aireleasetracker.com/model/moonshot/kimi-k3 - Developer: Moonshot AI - Release date: Thursday, Jul 16 2026 - Licensing: Open Weight - Parameters: 2.8T - Context window: 1M - Benchmarks published at release: - BullshitBench v2: 73% - DeepSWE 1.1: 69% - DeepSWE 1.0: 67.5% - Next.js Evals: 92% - Supabase Evals (with skills): 86.4% - Supabase Evals (no skills): 90.9% - Terminal-Bench 2.1: 88.3% - MCP Atlas: 84.2% - JobBench: 52.9% - Toolathlon-Verified: 73.2% - BrowseComp: 91.2% - Humanity's Last Exam (no tools): 43.5% - Humanity's Last Exam (with tools): 56% - GPQA Diamond: 93.5% - GDPval-AA v2: 1668 - CharXiv Reasoning: 84.8% - MMMU-Pro: 81.6% - Arena Elo (Text): 1486 - Arena Elo (Code): 1679 Kimi K3, released July 16, 2026, was Moonshot AI's biggest swing yet: a 2.8-trillion-parameter mixture-of-experts model — billed as the world's first open 3T-class release — activating 16 of 896 experts per token, with a 1M-token context window and native vision built on the lab's Kimi Delta Attention architecture. At launch it scored 93.5% on GPQA Diamond, the strongest open-weight result on that benchmark published at the time, alongside 88.3% on Terminal-Bench 2.1. Agentic work was the headline: 91.2% on BrowseComp — the best published score on this tracker at release — plus 56.0% on Humanity's Last Exam with tools and 84.2% on MCP Atlas. Moonshot's own framing placed K3 just behind Claude Fable 5 and GPT-5.6 Sol overall while undercutting both sharply on price, and the lab promised full open weights within days of the API launch — the same open-frontier playbook that began with K2 a year earlier, now at nearly triple the scale. **Q: When was Kimi K3 released?** A: Kimi K3 was released by Moonshot AI on Thursday, Jul 16 2026. **Q: Who made Kimi K3?** A: Kimi K3 was built by Moonshot AI. Chinese AI company building the Kimi family of models. Founded in 2023, known for long-context and efficient MoE architectures. **Q: What benchmark scores did Kimi K3 get?** A: Kimi K3 reports 19 tracked benchmark scores — BullshitBench v2: 73%; DeepSWE 1.1: 69%; DeepSWE 1.0: 67.5%; Next.js Evals: 92%; Supabase Evals (with skills): 86.4%; Supabase Evals (no skills): 90.9%; Terminal-Bench 2.1: 88.3%; MCP Atlas: 84.2%; JobBench: 52.9%; Toolathlon-Verified: 73.2%; BrowseComp: 91.2%; Humanity's Last Exam (no tools): 43.5%; Humanity's Last Exam (with tools): 56%; GPQA Diamond: 93.5%; GDPval-AA v2: 1668; CharXiv Reasoning: 84.8%; MMMU-Pro: 81.6%; Arena Elo (Text): 1486; Arena Elo (Code): 1679. Scores are the figures published at release by Moonshot AI. It holds the best score among all models tracked here on DeepSWE 1.0, BrowseComp and Arena Elo (Code). **Q: What is the context window of Kimi K3?** A: Kimi K3 has a context window of 1M. That is the maximum amount of input plus output the model can hold in a single request. **Q: How many parameters does Kimi K3 have?** A: Kimi K3 is reported at 2.8T parameters. **Q: Is Kimi K3 open source?** A: Partly. Kimi K3 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Kimi K3?** A: Moonshot AI's previous tracked release was Kimi K2.7 Code on Jun 12 2026, 34 days earlier. It is the most recent Moonshot AI model tracked on AI Release Tracker. ### Muse Spark 1.1 - URL: https://aireleasetracker.com/model/meta/muse-spark-1.1 - Developer: Meta - Release date: Thursday, Jul 9 2026 - Licensing: Proprietary - Benchmarks published at release: - SWE-Bench Pro: 61.5% - DeepSWE 1.1: 53.3% - Terminal-Bench 2.1: 80% - MCP Atlas: 88.1% - JobBench: 54.7% - Toolathlon-Verified: 75.6% - Humanity's Last Exam (with tools): 62.1% - OSWorld-Verified: 80.8% - Finance Agent v2: 57.2% - Harvey's Legal Agent Benchmark: 20% - TaxEval v2: 79.72% - MedScribe: 88.89% - CharXiv Reasoning: 88.4% - BabyVision: 76.3% - Arena Elo (Text): 1490 - Arena Elo (Code): 1540 **Q: When was Muse Spark 1.1 released?** A: Muse Spark 1.1 was released by Meta on Thursday, Jul 9 2026. **Q: Who made Muse Spark 1.1?** A: Muse Spark 1.1 was built by Meta. Develops the open-weight Llama series of models. Committed to open-source AI research. **Q: What benchmark scores did Muse Spark 1.1 get?** A: Muse Spark 1.1 reports 16 tracked benchmark scores — SWE-Bench Pro: 61.5%; DeepSWE 1.1: 53.3%; Terminal-Bench 2.1: 80%; MCP Atlas: 88.1%; JobBench: 54.7%; Toolathlon-Verified: 75.6%; Humanity's Last Exam (with tools): 62.1%; OSWorld-Verified: 80.8%; Finance Agent v2: 57.2%; Harvey's Legal Agent Benchmark: 20%; TaxEval v2: 79.72%; MedScribe: 88.89%; CharXiv Reasoning: 88.4%; BabyVision: 76.3%; Arena Elo (Text): 1490; Arena Elo (Code): 1540. Scores are the figures published at release by Meta. It holds the best score among all models tracked here on MCP Atlas, JobBench, Toolathlon-Verified, TaxEval v2 and MedScribe. **Q: Is Muse Spark 1.1 open source?** A: No. Muse Spark 1.1 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Muse Spark 1.1?** A: Meta's previous tracked release was Muse Spark on Apr 8 2026, 92 days earlier. It was followed by Muse Spark 1.2 on Aug 5 2026. ### Grok 4.5 - URL: https://aireleasetracker.com/model/xai/grok-4.5 - Developer: SpaceXAI - Release date: Wednesday, Jul 8 2026 - Licensing: Proprietary - Benchmarks published at release: - BullshitBench v2: 54% - Gray Swan IPI (k = 1): 13.4% - Gray Swan IPI (k = 10): 54.2% - Gray Swan IPI (k = 15): 60.8% - SWE-Bench Pro: 64.7% - SWE-Bench Multilingual: 78% - CursorBench v3.2: 66.7% - DeepSWE 1.1: 54% - DeepSWE 1.0: 62% - FrontierCode v1.1 (Extended) (extended split): 56.6% - APEX-SWE: 53.6% - Next.js Evals: 83% - Frontier-Bench v0.1: 17.8% - Terminal-Bench 3.0: 15.7% - Terminal-Bench 2.1: 83.3% - APEX-Agents: 47.1% - Harvey's Legal Agent Benchmark: 12.92% - MedScribe: 86.88% - AA Intelligence Index: 56 - GDPval-AA v2: 1526 - AA-Briefcase: 1313 - Arena Elo (Text): 1468 - Arena Elo (Code): 1549 Grok 4.5, released July 8, 2026, arrived as xAI's flagship and its strongest coding model at the time: a WebDev Arena Elo of 1566, 64.7% on SWE-Bench Pro, 78.0% on SWE-Bench Multilingual, and 83.3% on Terminal-Bench 2.1 — second only to Claude Fable 5 on the terminal-agent benchmark at release. It also posted 86.88% on MedScribe for clinical documentation and 62.0% on DeepSWE 1.0. Arriving almost exactly a year after Grok 4, it landed in the mid-2026 frontier cohort against Claude Fable 5, GPT-5.6, and Gemini 3.5. Its run as flagship was short: Grok 4.6 replaced it five weeks later, on August 12, 2026, using the same base model with heavier post-training. **Q: When was Grok 4.5 released?** A: Grok 4.5 was released by SpaceXAI on Wednesday, Jul 8 2026. **Q: Who made Grok 4.5?** A: Grok 4.5 was built by SpaceXAI. Elon Musk's AI company building the Grok series of models. Founded in 2023 as xAI, now part of SpaceX. **Q: What benchmark scores did Grok 4.5 get?** A: Grok 4.5 reports 23 tracked benchmark scores — BullshitBench v2: 54%; Gray Swan IPI (k = 1): 13.4%; Gray Swan IPI (k = 10): 54.2%; Gray Swan IPI (k = 15): 60.8%; SWE-Bench Pro: 64.7%; SWE-Bench Multilingual: 78%; CursorBench v3.2: 66.7%; DeepSWE 1.1: 54%; DeepSWE 1.0: 62%; FrontierCode v1.1 (Extended) (extended split): 56.6%; APEX-SWE: 53.6%; Next.js Evals: 83%; Frontier-Bench v0.1: 17.8%; Terminal-Bench 3.0: 15.7%; Terminal-Bench 2.1: 83.3%; APEX-Agents: 47.1%; Harvey's Legal Agent Benchmark: 12.92%; MedScribe: 86.88%; AA Intelligence Index: 56; GDPval-AA v2: 1526; AA-Briefcase: 1313; Arena Elo (Text): 1468; Arena Elo (Code): 1549. Scores are the figures published at release by SpaceXAI. **Q: Is Grok 4.5 open source?** A: No. Grok 4.5 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Grok 4.5?** A: SpaceXAI's previous tracked release was Grok 4.3 Beta on Apr 17 2026, 82 days earlier. It was followed by Grok 4.6 on Aug 12 2026. ### Claude Sonnet 5 - URL: https://aireleasetracker.com/model/anthropic/claude-sonnet-5 - Developer: Anthropic - Release date: Tuesday, Jun 30 2026 - Licensing: Proprietary - Benchmarks published at release: - BullshitBench v2: 80% - Gray Swan IPI (k = 1): 0.6% - Gray Swan IPI (k = 10): 4.7% - Gray Swan IPI (k = 15): 5.9% - SWE-Bench Pro: 63.2% - CursorBench v3.2: 61.5% - CursorBench v3.1: 61.2% - Next.js Evals: 79% - Supabase Evals (with skills): 95.5% - Supabase Evals (no skills): 90.9% - Frontier-Bench v0.1: 14.6% - Terminal-Bench 2.1: 80.4% - BrowseComp: 84.7% - Humanity's Last Exam (no tools): 43.2% - Humanity's Last Exam (with tools): 57.4% - OSWorld-Verified: 81.2% - GDPval-AA: 1618 - Arena Elo (Text): 1463 - Arena Elo (Code): 1543 **Q: When was Claude Sonnet 5 released?** A: Claude Sonnet 5 was released by Anthropic on Tuesday, Jun 30 2026. **Q: Who made Claude Sonnet 5?** A: Claude Sonnet 5 was built by Anthropic. AI safety company building the Claude family of models. Founded in 2021 by former OpenAI researchers. **Q: What benchmark scores did Claude Sonnet 5 get?** A: Claude Sonnet 5 reports 19 tracked benchmark scores — BullshitBench v2: 80%; Gray Swan IPI (k = 1): 0.6%; Gray Swan IPI (k = 10): 4.7%; Gray Swan IPI (k = 15): 5.9%; SWE-Bench Pro: 63.2%; CursorBench v3.2: 61.5%; CursorBench v3.1: 61.2%; Next.js Evals: 79%; Supabase Evals (with skills): 95.5%; Supabase Evals (no skills): 90.9%; Frontier-Bench v0.1: 14.6%; Terminal-Bench 2.1: 80.4%; BrowseComp: 84.7%; Humanity's Last Exam (no tools): 43.2%; Humanity's Last Exam (with tools): 57.4%; OSWorld-Verified: 81.2%; GDPval-AA: 1618; Arena Elo (Text): 1463; Arena Elo (Code): 1543. Scores are the figures published at release by Anthropic. **Q: Is Claude Sonnet 5 open source?** A: No. Claude Sonnet 5 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Claude Sonnet 5?** A: Anthropic's previous tracked release was Claude Fable 5 on Jun 9 2026, 21 days earlier. It was followed by Claude Opus 5 on Jul 24 2026. ### GPT-5.6 Sol - URL: https://aireleasetracker.com/model/openai/gpt-5.6-sol - Developer: OpenAI - Release date: Friday, Jun 26 2026 - Licensing: Proprietary - Benchmarks published at release: - BullshitBench v2: 47% - Gray Swan IPI (k = 1): 3.1% - Gray Swan IPI (k = 10): 16.3% - Gray Swan IPI (k = 15): 20% - CursorBench v3.2: 67.2% - DeepSWE 1.1: 73% - Next.js Evals: 92% - Supabase Evals (with skills): 95.5% - Supabase Evals (no skills): 90.9% - Frontier-Bench v0.1: 34.4% - Terminal-Bench 2.1: 88.8% - BU Bench: 67% - GDPval-AA v2: 1748 - Arena Elo (Text): 1486 - Arena Elo (Code): 1620 **Q: When was GPT-5.6 Sol released?** A: GPT-5.6 Sol was released by OpenAI on Friday, Jun 26 2026. **Q: Who made GPT-5.6 Sol?** A: GPT-5.6 Sol was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: What benchmark scores did GPT-5.6 Sol get?** A: GPT-5.6 Sol reports 15 tracked benchmark scores — BullshitBench v2: 47%; Gray Swan IPI (k = 1): 3.1%; Gray Swan IPI (k = 10): 16.3%; Gray Swan IPI (k = 15): 20%; CursorBench v3.2: 67.2%; DeepSWE 1.1: 73%; Next.js Evals: 92%; Supabase Evals (with skills): 95.5%; Supabase Evals (no skills): 90.9%; Frontier-Bench v0.1: 34.4%; Terminal-Bench 2.1: 88.8%; BU Bench: 67%; GDPval-AA v2: 1748; Arena Elo (Text): 1486; Arena Elo (Code): 1620. Scores are the figures published at release by OpenAI. It holds the best score among all models tracked here on DeepSWE 1.1, Supabase Evals (with skills), Supabase Evals (no skills) and Terminal-Bench 2.1. **Q: Is GPT-5.6 Sol open source?** A: No. GPT-5.6 Sol is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after GPT-5.6 Sol?** A: OpenAI's previous tracked release was GPT-5.5-Cyber on Jun 22 2026, 4 days earlier. It was followed by GPT-5.6 Terra on Jun 26 2026. ### GPT-5.6 Terra - URL: https://aireleasetracker.com/model/openai/gpt-5.6-terra - Developer: OpenAI - Release date: Friday, Jun 26 2026 - Licensing: Proprietary - Benchmarks published at release: - BullshitBench v2: 53% - Gray Swan IPI (k = 1): 5.4% - Gray Swan IPI (k = 10): 26% - Gray Swan IPI (k = 15): 30.4% - CursorBench v3.2: 64.9% - Frontier-Bench v0.1: 20.8% - Terminal-Bench 2.1: 84.3% - Arena Elo (Text): 1467 - Arena Elo (Code): 1526 **Q: When was GPT-5.6 Terra released?** A: GPT-5.6 Terra was released by OpenAI on Friday, Jun 26 2026. **Q: Who made GPT-5.6 Terra?** A: GPT-5.6 Terra was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: What benchmark scores did GPT-5.6 Terra get?** A: GPT-5.6 Terra reports 9 tracked benchmark scores — BullshitBench v2: 53%; Gray Swan IPI (k = 1): 5.4%; Gray Swan IPI (k = 10): 26%; Gray Swan IPI (k = 15): 30.4%; CursorBench v3.2: 64.9%; Frontier-Bench v0.1: 20.8%; Terminal-Bench 2.1: 84.3%; Arena Elo (Text): 1467; Arena Elo (Code): 1526. Scores are the figures published at release by OpenAI. **Q: Is GPT-5.6 Terra open source?** A: No. GPT-5.6 Terra is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after GPT-5.6 Terra?** A: OpenAI's previous tracked release was GPT-5.6 Sol on Jun 26 2026. It was followed by GPT-5.6 Luna on Jun 26 2026. ### GPT-5.6 Luna - URL: https://aireleasetracker.com/model/openai/gpt-5.6-luna - Developer: OpenAI - Release date: Friday, Jun 26 2026 - Licensing: Proprietary - Benchmarks published at release: - BullshitBench v2: 40% - Gray Swan IPI (k = 1): 8.3% - Gray Swan IPI (k = 10): 38.6% - Gray Swan IPI (k = 15): 43.9% - CursorBench v3.2: 61.1% - Frontier-Bench v0.1: 14.3% - Terminal-Bench 2.1: 82.5% - Arena Elo (Code): 1523 **Q: When was GPT-5.6 Luna released?** A: GPT-5.6 Luna was released by OpenAI on Friday, Jun 26 2026. **Q: Who made GPT-5.6 Luna?** A: GPT-5.6 Luna was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: What benchmark scores did GPT-5.6 Luna get?** A: GPT-5.6 Luna reports 8 tracked benchmark scores — BullshitBench v2: 40%; Gray Swan IPI (k = 1): 8.3%; Gray Swan IPI (k = 10): 38.6%; Gray Swan IPI (k = 15): 43.9%; CursorBench v3.2: 61.1%; Frontier-Bench v0.1: 14.3%; Terminal-Bench 2.1: 82.5%; Arena Elo (Code): 1523. Scores are the figures published at release by OpenAI. **Q: Is GPT-5.6 Luna open source?** A: No. GPT-5.6 Luna is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after GPT-5.6 Luna?** A: OpenAI's previous tracked release was GPT-5.6 Terra on Jun 26 2026. It was followed by GPT-5.6-Cyber on Aug 10 2026. ### GPT-5.5-Cyber - URL: https://aireleasetracker.com/model/openai/gpt-5.5-cyber - Developer: OpenAI - Release date: Monday, Jun 22 2026 - Licensing: Proprietary - Benchmarks published at release: - CyberGym: 85.6% GPT-5.5-Cyber, announced June 22, 2026, was OpenAI's first cybersecurity-specialised model and the anchor of Daybreak, the vetted-access program for security defenders that OpenAI had introduced a month earlier. Trained to find, validate, and help patch software vulnerabilities across large codebases, it scored 85.6% on CyberGym at release, against the 81.8% posted by the general-purpose GPT-5.5. The model never shipped to the public API — access ran through Daybreak's identity checks and monitoring, a template Google followed a month later with Gemini 3.5 Flash Cyber. Alongside the model, OpenAI and Trail of Bits launched Patch the Planet, which put the model to work on open-source projects including cURL and NATS Server. GPT-5.6-Cyber succeeded it in August 2026. **Q: When was GPT-5.5-Cyber released?** A: GPT-5.5-Cyber was released by OpenAI on Monday, Jun 22 2026. **Q: Who made GPT-5.5-Cyber?** A: GPT-5.5-Cyber was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: What benchmark scores did GPT-5.5-Cyber get?** A: GPT-5.5-Cyber reports 1 tracked benchmark score — CyberGym: 85.6%. Scores are the figures published at release by OpenAI. It holds the best score among all models tracked here on CyberGym. **Q: Is GPT-5.5-Cyber open source?** A: No. GPT-5.5-Cyber is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after GPT-5.5-Cyber?** A: OpenAI's previous tracked release was GPT-5.5-Pro on Apr 23 2026, 60 days earlier. It was followed by GPT-5.6 Sol on Jun 26 2026. ### GLM-5.2 - URL: https://aireleasetracker.com/model/zai/glm-5.2 - Developer: Z.ai - Release date: Tuesday, Jun 16 2026 - Licensing: Open Weight - Parameters: 744B - Context window: 1M - Benchmarks published at release: - BullshitBench v2: 31% - SWE-Bench Pro: 62.1% - CursorBench v3.2: 55% - CursorBench v3.1: 54.6% - DeepSWE 1.1: 44% - Next.js Evals: 88% - Frontier-Bench v0.1: 5.1% - Terminal-Bench 2.1: 81% - Humanity's Last Exam (no tools): 40.5% - Humanity's Last Exam (with tools): 54.7% - GPQA Diamond: 91.2% - GDPval-AA v2: 1514 - Arena Elo (Code): 1587 **Q: When was GLM-5.2 released?** A: GLM-5.2 was released by Z.ai on Tuesday, Jun 16 2026. **Q: Who made GLM-5.2?** A: GLM-5.2 was built by Z.ai. Chinese AI lab spun out of Tsinghua University (formerly Zhipu AI), building the open-weight GLM family. Rebranded internationally as Z.ai in 2025. **Q: What benchmark scores did GLM-5.2 get?** A: GLM-5.2 reports 13 tracked benchmark scores — BullshitBench v2: 31%; SWE-Bench Pro: 62.1%; CursorBench v3.2: 55%; CursorBench v3.1: 54.6%; DeepSWE 1.1: 44%; Next.js Evals: 88%; Frontier-Bench v0.1: 5.1%; Terminal-Bench 2.1: 81%; Humanity's Last Exam (no tools): 40.5%; Humanity's Last Exam (with tools): 54.7%; GPQA Diamond: 91.2%; GDPval-AA v2: 1514; Arena Elo (Code): 1587. Scores are the figures published at release by Z.ai. **Q: What is the context window of GLM-5.2?** A: GLM-5.2 has a context window of 1M. That is the maximum amount of input plus output the model can hold in a single request. **Q: How many parameters does GLM-5.2 have?** A: GLM-5.2 is reported at 744B parameters. **Q: Is GLM-5.2 open source?** A: Partly. GLM-5.2 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after GLM-5.2?** A: Z.ai's previous tracked release was GLM-5.1 on Apr 7 2026, 70 days earlier. It was followed by GLM-5.3 on Aug 14 2026. ### Kimi K2.7 Code - URL: https://aireleasetracker.com/model/moonshot/kimi-k2.7-code - Developer: Moonshot AI - Release date: Friday, Jun 12 2026 - Licensing: Open Weight - Parameters: 1T - Context window: 256k - Benchmarks published at release: - CursorBench v3.2: 49.7% - DeepSWE 1.1: 31% - Next.js Evals: 75% - Arena Elo (Code): 1469 **Q: When was Kimi K2.7 Code released?** A: Kimi K2.7 Code was released by Moonshot AI on Friday, Jun 12 2026. **Q: Who made Kimi K2.7 Code?** A: Kimi K2.7 Code was built by Moonshot AI. Chinese AI company building the Kimi family of models. Founded in 2023, known for long-context and efficient MoE architectures. **Q: What benchmark scores did Kimi K2.7 Code get?** A: Kimi K2.7 Code reports 4 tracked benchmark scores — CursorBench v3.2: 49.7%; DeepSWE 1.1: 31%; Next.js Evals: 75%; Arena Elo (Code): 1469. Scores are the figures published at release by Moonshot AI. **Q: What is the context window of Kimi K2.7 Code?** A: Kimi K2.7 Code has a context window of 256k. That is the maximum amount of input plus output the model can hold in a single request. **Q: How many parameters does Kimi K2.7 Code have?** A: Kimi K2.7 Code is reported at 1T parameters. **Q: Is Kimi K2.7 Code open source?** A: Partly. Kimi K2.7 Code is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Kimi K2.7 Code?** A: Moonshot AI's previous tracked release was Kimi K2.6 on Apr 21 2026, 52 days earlier. It was followed by Kimi K3 on Jul 16 2026. ### Claude Fable 5 - URL: https://aireleasetracker.com/model/anthropic/claude-fable-5 - Developer: Anthropic - Release date: Tuesday, Jun 9 2026 - Licensing: Proprietary - Context window: 1M - Benchmarks published at release: - BullshitBench v2: 54% - Gray Swan IPI (k = 1): 0.4% - Gray Swan IPI (k = 10): 2.3% - Gray Swan IPI (k = 15): 2.8% - SWE-Bench Pro: 80.3% - SWE-Bench Verified: 95.5% - CursorBench v3.2: 70.5% - CursorBench v3.1: 72.9% - DeepSWE 1.1: 70% - DeepSWE 1.0: 66.1% - Next.js Evals: 92% - Frontier-Bench v0.1: 33.8% - Terminal-Bench 2.1: 88% - BrowseComp: 86.9% - Humanity's Last Exam (with tools): 64.5% - OSWorld-Verified: 85% - Harvey's Legal Agent Benchmark: 11.25% - TaxEval v2: 76.94% - MedScribe: 88.52% - GDPval-AA: 1932 - GDPval-AA v2: 1760 - Arena Elo (Text): 1509 - Arena Elo (Code): 1631 Claude Fable 5, released June 9, 2026, is the first model of Anthropic's Claude 5 generation and the debut of the Mythos-class tier that sits above Opus in the lineup. It posted the largest single-release jump in the tracker's dataset: 95.5% on SWE-Bench Verified, 80.3% on SWE-Bench Pro, 88.0% on Terminal-Bench 2.1, and a WebDev Arena coding Elo of 1649 — over 80 points clear of the next model at launch. Fable 5 ships with a 1M-token context window and leads agentic evaluations including OSWorld-Verified (85.0%) and Artificial Analysis GDPval (1932 Elo). Anthropic offers the same underlying model without additional dual-use safety measures as Claude Mythos 5 to approved organisations. Claude Sonnet 5 followed on June 30, 2026, and Claude Opus 5 brought the 5-series to the Opus tier on July 24, 2026 at half Fable's price. **Q: When was Claude Fable 5 released?** A: Claude Fable 5 was released by Anthropic on Tuesday, Jun 9 2026. **Q: Who made Claude Fable 5?** A: Claude Fable 5 was built by Anthropic. AI safety company building the Claude family of models. Founded in 2021 by former OpenAI researchers. **Q: What benchmark scores did Claude Fable 5 get?** A: Claude Fable 5 reports 23 tracked benchmark scores — BullshitBench v2: 54%; Gray Swan IPI (k = 1): 0.4%; Gray Swan IPI (k = 10): 2.3%; Gray Swan IPI (k = 15): 2.8%; SWE-Bench Pro: 80.3%; SWE-Bench Verified: 95.5%; CursorBench v3.2: 70.5%; CursorBench v3.1: 72.9%; DeepSWE 1.1: 70%; DeepSWE 1.0: 66.1%; Next.js Evals: 92%; Frontier-Bench v0.1: 33.8%; Terminal-Bench 2.1: 88%; BrowseComp: 86.9%; Humanity's Last Exam (with tools): 64.5%; OSWorld-Verified: 85%; Harvey's Legal Agent Benchmark: 11.25%; TaxEval v2: 76.94%; MedScribe: 88.52%; GDPval-AA: 1932; GDPval-AA v2: 1760; Arena Elo (Text): 1509; Arena Elo (Code): 1631. Scores are the figures published at release by Anthropic. It holds the best score among all models tracked here on SWE-Bench Pro, SWE-Bench Verified, CursorBench v3.2, CursorBench v3.1, GDPval-AA and Arena Elo (Text). **Q: What is the context window of Claude Fable 5?** A: Claude Fable 5 has a context window of 1M. That is the maximum amount of input plus output the model can hold in a single request. **Q: Is Claude Fable 5 open source?** A: No. Claude Fable 5 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Claude Fable 5?** A: Anthropic's previous tracked release was Claude Opus 4.8 on May 28 2026, 12 days earlier. It was followed by Claude Sonnet 5 on Jun 30 2026. ### Qwen3.7-Plus - URL: https://aireleasetracker.com/model/qwen/qwen3.7-plus - Developer: Qwen - Release date: Monday, Jun 1 2026 - Licensing: Proprietary - Benchmarks published at release: - LiveCodeBench: 89.6% - Arena Elo (Text): 1458 **Q: When was Qwen3.7-Plus released?** A: Qwen3.7-Plus was released by Qwen on Monday, Jun 1 2026. **Q: Who made Qwen3.7-Plus?** A: Qwen3.7-Plus was built by Qwen. Alibaba's AI lab, building the Qwen family. The most prolific publisher of open-weight models of any major lab, alongside a proprietary Max and Plus tier sold through Alibaba Cloud. **Q: What benchmark scores did Qwen3.7-Plus get?** A: Qwen3.7-Plus reports 2 tracked benchmark scores — LiveCodeBench: 89.6%; Arena Elo (Text): 1458. Scores are the figures published at release by Qwen. **Q: Is Qwen3.7-Plus open source?** A: No. Qwen3.7-Plus is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Qwen3.7-Plus?** A: Qwen's previous tracked release was Qwen3.7-Max on May 20 2026, 12 days earlier. It was followed by Qwen3.8-Max on Aug 3 2026. ### Claude Opus 4.8 - URL: https://aireleasetracker.com/model/anthropic/claude-opus-4.8 - Developer: Anthropic - Release date: Thursday, May 28 2026 - Licensing: Proprietary - Context window: 1M - Benchmarks published at release: - BullshitBench v2: 95% - Gray Swan IPI (k = 1): 0.5% - Gray Swan IPI (k = 10): 4.1% - Gray Swan IPI (k = 15): 5.5% - SWE-Bench Pro: 69.2% - SWE-Bench Multilingual: 84.4% - CursorBench v3.2: 62.3% - CursorBench v3.1: 63.8% - DeepSWE 1.0: 55.8% - Next.js Evals: 88% - Frontier-Bench v0.1: 21.1% - Terminal-Bench 2.1: 74.6% - BU Bench: 74% - BrowseComp: 84.3% - Humanity's Last Exam (no tools): 49.8% - Humanity's Last Exam (with tools): 57.9% - OSWorld-Verified: 83.4% - Finance Agent v2: 53.9% - Harvey's Legal Agent Benchmark: 9.58% - TaxEval v2: 75.63% - MedScribe: 85.75% - GDPval-AA: 1890 - GDPval-AA v2: 1600 - Arena Elo (Text): 1482 - Arena Elo (Code): 1568 Claude Opus 4.8, released May 28, 2026, continued the rapid iteration of the Opus line with a 1M-token context window and top scores across agentic and real-world evaluations: 69.2% on SWE-Bench Pro, 84.4% on SWE-Bench Multilingual, and an Artificial Analysis GDPval Elo of 1890, the highest of any model at its release. Its 95% score on BullshitBench-V2 — a benchmark measuring whether a model pushes back on false premises — was also a record. Opus 4.8 arrived just seven weeks after Opus 4.7 and held the flagship spot for less than two weeks before Claude Fable 5, the first of Anthropic's Claude 5 generation, launched on June 9, 2026. It also closed out the 4-series for the Opus name: Claude Opus 5 succeeded it on July 24, 2026 at the same price. **Q: When was Claude Opus 4.8 released?** A: Claude Opus 4.8 was released by Anthropic on Thursday, May 28 2026. **Q: Who made Claude Opus 4.8?** A: Claude Opus 4.8 was built by Anthropic. AI safety company building the Claude family of models. Founded in 2021 by former OpenAI researchers. **Q: What benchmark scores did Claude Opus 4.8 get?** A: Claude Opus 4.8 reports 25 tracked benchmark scores — BullshitBench v2: 95%; Gray Swan IPI (k = 1): 0.5%; Gray Swan IPI (k = 10): 4.1%; Gray Swan IPI (k = 15): 5.5%; SWE-Bench Pro: 69.2%; SWE-Bench Multilingual: 84.4%; CursorBench v3.2: 62.3%; CursorBench v3.1: 63.8%; DeepSWE 1.0: 55.8%; Next.js Evals: 88%; Frontier-Bench v0.1: 21.1%; Terminal-Bench 2.1: 74.6%; BU Bench: 74%; BrowseComp: 84.3%; Humanity's Last Exam (no tools): 49.8%; Humanity's Last Exam (with tools): 57.9%; OSWorld-Verified: 83.4%; Finance Agent v2: 53.9%; Harvey's Legal Agent Benchmark: 9.58%; TaxEval v2: 75.63%; MedScribe: 85.75%; GDPval-AA: 1890; GDPval-AA v2: 1600; Arena Elo (Text): 1482; Arena Elo (Code): 1568. Scores are the figures published at release by Anthropic. It holds the best score among all models tracked here on BullshitBench v2, SWE-Bench Multilingual and BU Bench. **Q: What is the context window of Claude Opus 4.8?** A: Claude Opus 4.8 has a context window of 1M. That is the maximum amount of input plus output the model can hold in a single request. **Q: Is Claude Opus 4.8 open source?** A: No. Claude Opus 4.8 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Claude Opus 4.8?** A: Anthropic's previous tracked release was Claude Opus 4.7 on Apr 16 2026, 42 days earlier. It was followed by Claude Fable 5 on Jun 9 2026. ### Qwen3.7-Max - URL: https://aireleasetracker.com/model/qwen/qwen3.7-max - Developer: Qwen - Release date: Wednesday, May 20 2026 - Licensing: Proprietary - Benchmarks published at release: - BullshitBench v2: 71% - SWE-Bench Pro: 60.6% - LiveCodeBench: 91.6% - Terminal-Bench 2.0: 69.7% - MCP Atlas: 76.4% - GPQA Diamond: 92.4% - Arena Elo (Text): 1475 **Q: When was Qwen3.7-Max released?** A: Qwen3.7-Max was released by Qwen on Wednesday, May 20 2026. **Q: Who made Qwen3.7-Max?** A: Qwen3.7-Max was built by Qwen. Alibaba's AI lab, building the Qwen family. The most prolific publisher of open-weight models of any major lab, alongside a proprietary Max and Plus tier sold through Alibaba Cloud. **Q: What benchmark scores did Qwen3.7-Max get?** A: Qwen3.7-Max reports 7 tracked benchmark scores — BullshitBench v2: 71%; SWE-Bench Pro: 60.6%; LiveCodeBench: 91.6%; Terminal-Bench 2.0: 69.7%; MCP Atlas: 76.4%; GPQA Diamond: 92.4%; Arena Elo (Text): 1475. Scores are the figures published at release by Qwen. **Q: Is Qwen3.7-Max open source?** A: No. Qwen3.7-Max is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Qwen3.7-Max?** A: Qwen's previous tracked release was Qwen3.6 on Apr 16 2026, 34 days earlier. It was followed by Qwen3.7-Plus on Jun 1 2026. ### Gemini 3.5 Flash - URL: https://aireleasetracker.com/model/google/gemini-3.5-flash - Developer: Google - Release date: Tuesday, May 19 2026 - Licensing: Proprietary - Benchmarks published at release: - BullshitBench v2: 20% - Gray Swan IPI (k = 1): 14.1% - Gray Swan IPI (k = 10): 54.2% - Gray Swan IPI (k = 15): 60.5% - SWE-Bench Pro: 55.1% - CursorBench v3.2: 48.8% - CursorBench v3.1: 49.8% - DeepSWE 1.1: 37% - MLE-Bench: 49.7% - Terminal-Bench 2.1: 76.2% - MCP Atlas: 83.6% - Toolathlon: 56.5% - BU Bench: 58% - Humanity's Last Exam (no tools): 40.2% - ARC-AGI-2: 72.1% - OSWorld-Verified: 78.4% - Finance Agent v2: 57.9% - GDPval-AA: 1656 - GDPval-AA v2: 1349 - CharXiv Reasoning: 84.2% - MMMU-Pro: 83.6% - Blueprint-Bench 2: 33.6% - MRCR v2 (8-needle) (128k average): 77.3% - MRCR v2 (8-needle) (1M pointwise): 26.6% - Arena Elo (Text): 1476 - Arena Elo (Code): 1509 **Q: When was Gemini 3.5 Flash released?** A: Gemini 3.5 Flash was released by Google on Tuesday, May 19 2026. **Q: Who made Gemini 3.5 Flash?** A: Gemini 3.5 Flash was built by Google. Builds the Gemini family of models through Google DeepMind. Integrates AI across Google products. **Q: What benchmark scores did Gemini 3.5 Flash get?** A: Gemini 3.5 Flash reports 26 tracked benchmark scores — BullshitBench v2: 20%; Gray Swan IPI (k = 1): 14.1%; Gray Swan IPI (k = 10): 54.2%; Gray Swan IPI (k = 15): 60.5%; SWE-Bench Pro: 55.1%; CursorBench v3.2: 48.8%; CursorBench v3.1: 49.8%; DeepSWE 1.1: 37%; MLE-Bench: 49.7%; Terminal-Bench 2.1: 76.2%; MCP Atlas: 83.6%; Toolathlon: 56.5%; BU Bench: 58%; Humanity's Last Exam (no tools): 40.2%; ARC-AGI-2: 72.1%; OSWorld-Verified: 78.4%; Finance Agent v2: 57.9%; GDPval-AA: 1656; GDPval-AA v2: 1349; CharXiv Reasoning: 84.2%; MMMU-Pro: 83.6%; Blueprint-Bench 2: 33.6%; MRCR v2 (8-needle) (128k average): 77.3%; MRCR v2 (8-needle) (1M pointwise): 26.6%; Arena Elo (Text): 1476; Arena Elo (Code): 1509. Scores are the figures published at release by Google. It holds the best score among all models tracked here on Toolathlon, Finance Agent v2 and MMMU-Pro. **Q: Is Gemini 3.5 Flash open source?** A: No. Gemini 3.5 Flash is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Gemini 3.5 Flash?** A: Google's previous tracked release was Gemma 4 on Apr 2 2026, 47 days earlier. It was followed by Gemini 3.6 Flash on Jul 21 2026. ### Composer 2.5 - URL: https://aireleasetracker.com/model/cursor/composer-2.5 - Developer: Cursor - Release date: Monday, May 18 2026 - Licensing: Proprietary - Benchmarks published at release: - SWE-Bench Pro: 54% - SWE-Bench Multilingual: 79.8% - CursorBench v3.2: 56.1% - CursorBench v3.1: 63.2% - DeepSWE 1.0: 18% - Next.js Evals: 92% - Terminal-Bench 2.1: 73% - Terminal-Bench 2.0: 69.3% **Q: When was Composer 2.5 released?** A: Composer 2.5 was released by Cursor on Monday, May 18 2026. **Q: Who made Composer 2.5?** A: Composer 2.5 was built by Cursor. **Q: What benchmark scores did Composer 2.5 get?** A: Composer 2.5 reports 8 tracked benchmark scores — SWE-Bench Pro: 54%; SWE-Bench Multilingual: 79.8%; CursorBench v3.2: 56.1%; CursorBench v3.1: 63.2%; DeepSWE 1.0: 18%; Next.js Evals: 92%; Terminal-Bench 2.1: 73%; Terminal-Bench 2.0: 69.3%. Scores are the figures published at release by Cursor. It holds the best score among all models tracked here on Next.js Evals. **Q: Is Composer 2.5 open source?** A: No. Composer 2.5 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Composer 2.5?** A: Cursor's previous tracked release was Composer 2 on Mar 19 2026, 60 days earlier. It is the most recent Cursor model tracked on AI Release Tracker. ### Mistral Medium 3.5 - URL: https://aireleasetracker.com/model/mistral/mistral-medium-3.5 - Developer: Mistral - Release date: Wednesday, Apr 29 2026 - Licensing: Open Weight - Parameters: 128B - Context window: 256k - Benchmarks published at release: - SWE-Bench Verified: 77.6% - Arena Elo (Code): 1267 **Q: When was Mistral Medium 3.5 released?** A: Mistral Medium 3.5 was released by Mistral on Wednesday, Apr 29 2026. **Q: Who made Mistral Medium 3.5?** A: Mistral Medium 3.5 was built by Mistral. French AI company building open and commercial models. Founded in 2023 by former Meta and DeepMind researchers. **Q: What benchmark scores did Mistral Medium 3.5 get?** A: Mistral Medium 3.5 reports 2 tracked benchmark scores — SWE-Bench Verified: 77.6%; Arena Elo (Code): 1267. Scores are the figures published at release by Mistral. **Q: What is the context window of Mistral Medium 3.5?** A: Mistral Medium 3.5 has a context window of 256k. That is the maximum amount of input plus output the model can hold in a single request. **Q: How many parameters does Mistral Medium 3.5 have?** A: Mistral Medium 3.5 is reported at 128B parameters. **Q: Is Mistral Medium 3.5 open source?** A: Partly. Mistral Medium 3.5 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Mistral Medium 3.5?** A: Mistral's previous tracked release was Mistral Small 4 on Mar 16 2026, 44 days earlier. It is the most recent Mistral model tracked on AI Release Tracker. ### DeepSeek-V4-Pro - URL: https://aireleasetracker.com/model/deepseek/deepseek-v4-pro - Developer: DeepSeek - Release date: Friday, Apr 24 2026 - Licensing: Open Weight - Benchmarks published at release: - BullshitBench v2: 14% - LiveCodeBench: 93.5% - BrowseComp: 83.4% - GPQA Diamond: 90.1% - Arena Elo (Text): 1457 - Arena Elo (Code): 1446 DeepSeek-V4-Pro, released April 24, 2026 alongside the lighter DeepSeek-V4-Flash, launched as DeepSeek's frontier open-weight flagship. It scored 90.1% on GPQA Diamond and 93.5% on LiveCodeBench — the best published competitive-coding score on this tracker at the time — with 83.4% on BrowseComp for agentic web research. The V4 generation extended DeepSeek's consistent playbook: frontier-adjacent capability, open weights, and dramatically lower cost than closed rivals. At launch V4-Pro competed directly with the strongest proprietary models of spring 2026 while remaining freely downloadable, continuing the open-weights pressure the lab began with V3 and R1. DeepSeek later designated this build V4-Pro-Preview; the generation's next update, in July 2026, went to the Flash tier first, and the Pro tier was refreshed as V4-Pro-0813 in August 2026. **Q: When was DeepSeek-V4-Pro released?** A: DeepSeek-V4-Pro was released by DeepSeek on Friday, Apr 24 2026. **Q: Who made DeepSeek-V4-Pro?** A: DeepSeek-V4-Pro was built by DeepSeek. Chinese AI lab known for efficient, open-weight models. Gained attention for strong performance at lower cost. **Q: What benchmark scores did DeepSeek-V4-Pro get?** A: DeepSeek-V4-Pro reports 6 tracked benchmark scores — BullshitBench v2: 14%; LiveCodeBench: 93.5%; BrowseComp: 83.4%; GPQA Diamond: 90.1%; Arena Elo (Text): 1457; Arena Elo (Code): 1446. Scores are the figures published at release by DeepSeek. It holds the best score among all models tracked here on LiveCodeBench. **Q: Is DeepSeek-V4-Pro open source?** A: Partly. DeepSeek-V4-Pro is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after DeepSeek-V4-Pro?** A: DeepSeek's previous tracked release was DeepSeek-V3.2 on Dec 1 2025, 144 days earlier. It was followed by DeepSeek-V4-Flash on Apr 24 2026. ### DeepSeek-V4-Flash - URL: https://aireleasetracker.com/model/deepseek/deepseek-v4-flash - Developer: DeepSeek - Release date: Friday, Apr 24 2026 - Licensing: Open Weight - Benchmarks published at release: - BullshitBench v2: 18% - LiveCodeBench: 91.6% - GPQA Diamond: 88.1% - Arena Elo (Code): 1577 **Q: When was DeepSeek-V4-Flash released?** A: DeepSeek-V4-Flash was released by DeepSeek on Friday, Apr 24 2026. **Q: Who made DeepSeek-V4-Flash?** A: DeepSeek-V4-Flash was built by DeepSeek. Chinese AI lab known for efficient, open-weight models. Gained attention for strong performance at lower cost. **Q: What benchmark scores did DeepSeek-V4-Flash get?** A: DeepSeek-V4-Flash reports 4 tracked benchmark scores — BullshitBench v2: 18%; LiveCodeBench: 91.6%; GPQA Diamond: 88.1%; Arena Elo (Code): 1577. Scores are the figures published at release by DeepSeek. **Q: Is DeepSeek-V4-Flash open source?** A: Partly. DeepSeek-V4-Flash is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after DeepSeek-V4-Flash?** A: DeepSeek's previous tracked release was DeepSeek-V4-Pro on Apr 24 2026. It was followed by DeepSeek-V4-Flash-0731 on Jul 31 2026. ### GPT-5.5 - URL: https://aireleasetracker.com/model/openai/gpt-5.5 - Developer: OpenAI - Release date: Thursday, Apr 23 2026 - Licensing: Proprietary - Context window: 1.05M - Benchmarks published at release: - BullshitBench v2: 47% - Gray Swan IPI (k = 1): 3% - Gray Swan IPI (k = 10): 17.4% - Gray Swan IPI (k = 15): 20.8% - SWE-Bench Pro: 58.6% - SWE-Bench Multilingual: 77.8% - CursorBench v3.2: 58.4% - CursorBench v3.1: 64.3% - DeepSWE 1.0: 64.3% - Terminal-Bench 2.1: 78.2% - Terminal-Bench 2.0: 82.7% - Expert-SWE (Internal): 73.1% - MCP Atlas: 75.3% - Toolathlon: 55.6% - BrowseComp: 84.4% - CyberGym: 81.8% - Humanity's Last Exam (no tools): 41.4% - Humanity's Last Exam (with tools): 52.2% - ARC-AGI-2: 84.6% - FrontierMath (Tier 1–3): 51.7% - FrontierMath (Tier 4): 35.4% - GPQA Diamond: 93.6% - OSWorld-Verified: 78.7% - Finance Agent v2: 51.8% - Harvey's Legal Agent Benchmark: 3.75% - TaxEval v2: 74.98% - MedScribe: 86.87% - GDPval-AA: 1769 - GDPval-AA v2: 1494 - GDPval (win/tie rate): 84.9% - CharXiv Reasoning: 84.1% - MMMU-Pro: 81.2% - Blueprint-Bench 2: 36.2% - MRCR v2 (8-needle) (128k average): 94.8% - Arena Elo (Text): 1481 - Arena Elo (Code): 1509 GPT-5.5, released April 23, 2026 alongside GPT-5.5-Pro, pushed OpenAI's context window past the million-token mark (1.05M) and posted the strongest long-context recall in its class — 94.8% on MRCR v2 at 128K. Its 84.6% on ARC-AGI-2 and 51.7% on FrontierMath Tiers 1–3 led all models at release on abstract reasoning and research mathematics. On agentic work it scored 78.2% on Terminal-Bench 2.1, 78.7% on OSWorld-Verified, and 84.9% on GDPval win-rate — the highest real-world-task result of any model at the time. GPT-5.5 headlined OpenAI's lineup for two months until the three-model GPT-5.6 family (Sol, Terra, Luna) arrived on June 26, 2026. **Q: When was GPT-5.5 released?** A: GPT-5.5 was released by OpenAI on Thursday, Apr 23 2026. **Q: Who made GPT-5.5?** A: GPT-5.5 was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: What benchmark scores did GPT-5.5 get?** A: GPT-5.5 reports 36 tracked benchmark scores — BullshitBench v2: 47%; Gray Swan IPI (k = 1): 3%; Gray Swan IPI (k = 10): 17.4%; Gray Swan IPI (k = 15): 20.8%; SWE-Bench Pro: 58.6%; SWE-Bench Multilingual: 77.8%; CursorBench v3.2: 58.4%; CursorBench v3.1: 64.3%; DeepSWE 1.0: 64.3%; Terminal-Bench 2.1: 78.2%; Terminal-Bench 2.0: 82.7%; Expert-SWE (Internal): 73.1%; MCP Atlas: 75.3%; Toolathlon: 55.6%; BrowseComp: 84.4%; CyberGym: 81.8%; Humanity's Last Exam (no tools): 41.4%; Humanity's Last Exam (with tools): 52.2%; ARC-AGI-2: 84.6%; FrontierMath (Tier 1–3): 51.7%; FrontierMath (Tier 4): 35.4%; GPQA Diamond: 93.6%; OSWorld-Verified: 78.7%; Finance Agent v2: 51.8%; Harvey's Legal Agent Benchmark: 3.75%; TaxEval v2: 74.98%; MedScribe: 86.87%; GDPval-AA: 1769; GDPval-AA v2: 1494; GDPval (win/tie rate): 84.9%; CharXiv Reasoning: 84.1%; MMMU-Pro: 81.2%; Blueprint-Bench 2: 36.2%; MRCR v2 (8-needle) (128k average): 94.8%; Arena Elo (Text): 1481; Arena Elo (Code): 1509. Scores are the figures published at release by OpenAI. It holds the best score among all models tracked here on Terminal-Bench 2.0, Expert-SWE (Internal), ARC-AGI-2, GDPval (win/tie rate) and Blueprint-Bench 2. **Q: What is the context window of GPT-5.5?** A: GPT-5.5 has a context window of 1.05M. That is the maximum amount of input plus output the model can hold in a single request. **Q: Is GPT-5.5 open source?** A: No. GPT-5.5 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after GPT-5.5?** A: OpenAI's previous tracked release was GPT-5.4 nano on Mar 17 2026, 37 days earlier. It was followed by GPT-5.5-Pro on Apr 23 2026. ### GPT-5.5-Pro - URL: https://aireleasetracker.com/model/openai/gpt-5.5-pro - Developer: OpenAI - Release date: Thursday, Apr 23 2026 - Licensing: Proprietary - Benchmarks published at release: - BullshitBench v2: 36% - Next.js Evals: 83% - BrowseComp: 90.1% - Humanity's Last Exam (with tools): 57.2% - FrontierMath (Tier 1–3): 52.4% - FrontierMath (Tier 4): 39.6% - GDPval (win/tie rate): 82.3% **Q: When was GPT-5.5-Pro released?** A: GPT-5.5-Pro was released by OpenAI on Thursday, Apr 23 2026. **Q: Who made GPT-5.5-Pro?** A: GPT-5.5-Pro was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: What benchmark scores did GPT-5.5-Pro get?** A: GPT-5.5-Pro reports 7 tracked benchmark scores — BullshitBench v2: 36%; Next.js Evals: 83%; BrowseComp: 90.1%; Humanity's Last Exam (with tools): 57.2%; FrontierMath (Tier 1–3): 52.4%; FrontierMath (Tier 4): 39.6%; GDPval (win/tie rate): 82.3%. Scores are the figures published at release by OpenAI. It holds the best score among all models tracked here on FrontierMath (Tier 1–3) and FrontierMath (Tier 4). **Q: Is GPT-5.5-Pro open source?** A: No. GPT-5.5-Pro is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after GPT-5.5-Pro?** A: OpenAI's previous tracked release was GPT-5.5 on Apr 23 2026. It was followed by GPT-5.5-Cyber on Jun 22 2026. ### Kimi K2.6 - URL: https://aireleasetracker.com/model/moonshot/kimi-k2.6 - Developer: Moonshot AI - Release date: Tuesday, Apr 21 2026 - Licensing: Open Weight - Parameters: 1T - Context window: 256k - Benchmarks published at release: - BullshitBench v2: 65% - SWE-Bench Verified: 80.2% - CursorBench v3.1: 47.6% - Next.js Evals: 67% - LiveCodeBench: 89.6% - GPQA Diamond: 90.5% - Arena Elo (Text): 1461 - Arena Elo (Code): 1513 Kimi K2.6, released April 21, 2026, was Moonshot AI's trillion-parameter open-weight flagship of spring 2026, scoring 90.5% on GPQA Diamond and 80.2% on SWE-Bench Verified — numbers that matched or beat several proprietary frontier models of the same season — along with 89.6% on LiveCodeBench and a WebDev Arena Elo of 1513. The jump over January's K2.5 was substantial (up nearly three points on GPQA and 3.4 on SWE-Bench Verified in twelve weeks), keeping Moonshot in the small group of labs shipping frontier-class open weights. The coding-specialised Kimi K2.7 Code followed in June 2026, and the K2 era closed when the 2.8-trillion-parameter Kimi K3 arrived that July. **Q: When was Kimi K2.6 released?** A: Kimi K2.6 was released by Moonshot AI on Tuesday, Apr 21 2026. **Q: Who made Kimi K2.6?** A: Kimi K2.6 was built by Moonshot AI. Chinese AI company building the Kimi family of models. Founded in 2023, known for long-context and efficient MoE architectures. **Q: What benchmark scores did Kimi K2.6 get?** A: Kimi K2.6 reports 8 tracked benchmark scores — BullshitBench v2: 65%; SWE-Bench Verified: 80.2%; CursorBench v3.1: 47.6%; Next.js Evals: 67%; LiveCodeBench: 89.6%; GPQA Diamond: 90.5%; Arena Elo (Text): 1461; Arena Elo (Code): 1513. Scores are the figures published at release by Moonshot AI. **Q: What is the context window of Kimi K2.6?** A: Kimi K2.6 has a context window of 256k. That is the maximum amount of input plus output the model can hold in a single request. **Q: How many parameters does Kimi K2.6 have?** A: Kimi K2.6 is reported at 1T parameters. **Q: Is Kimi K2.6 open source?** A: Partly. Kimi K2.6 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Kimi K2.6?** A: Moonshot AI's previous tracked release was Kimi K2.5 on Jan 27 2026, 84 days earlier. It was followed by Kimi K2.7 Code on Jun 12 2026. ### Grok 4.3 Beta - URL: https://aireleasetracker.com/model/xai/grok-4.3-beta - Developer: SpaceXAI - Release date: Friday, Apr 17 2026 - Licensing: Proprietary - Benchmarks published at release: - BullshitBench v2: 50% **Q: When was Grok 4.3 Beta released?** A: Grok 4.3 Beta was released by SpaceXAI on Friday, Apr 17 2026. **Q: Who made Grok 4.3 Beta?** A: Grok 4.3 Beta was built by SpaceXAI. Elon Musk's AI company building the Grok series of models. Founded in 2023 as xAI, now part of SpaceX. **Q: What benchmark scores did Grok 4.3 Beta get?** A: Grok 4.3 Beta reports 1 tracked benchmark score — BullshitBench v2: 50%. Scores are the figures published at release by SpaceXAI. **Q: Is Grok 4.3 Beta open source?** A: No. Grok 4.3 Beta is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Grok 4.3 Beta?** A: SpaceXAI's previous tracked release was Grok 4.20 Beta on Feb 17 2026, 59 days earlier. It was followed by Grok 4.5 on Jul 8 2026. ### Claude Opus 4.7 - URL: https://aireleasetracker.com/model/anthropic/claude-opus-4.7 - Developer: Anthropic - Release date: Thursday, Apr 16 2026 - Licensing: Proprietary - Context window: 1M - Benchmarks published at release: - BullshitBench v2: 83% - SWE-Bench Pro: 64.3% - SWE-Bench Verified: 87.6% - SWE-Bench Multilingual: 80.5% - CursorBench v3.1: 64.8% - Next.js Evals: 75% - Terminal-Bench 2.1: 66.1% - Terminal-Bench 2.0: 69.4% - MCP Atlas: 79.1% - BrowseComp: 79.3% - CyberGym: 73.1% - Humanity's Last Exam (no tools): 46.9% - Humanity's Last Exam (with tools): 54.7% - ARC-AGI-2: 75.8% - FrontierMath (Tier 1–3): 43.8% - FrontierMath (Tier 4): 22.9% - GPQA Diamond: 94.2% - OSWorld-Verified: 78% - Finance Agent v2: 51.5% - GDPval-AA: 1753 - GDPval (win/tie rate): 80.3% - CharXiv Reasoning: 82.1% - MMMU-Pro: 75.2% - Blueprint-Bench 2: 24.5% - MRCR v2 (8-needle) (128k average): 59.3% - Arena Elo (Text): 1503 - Arena Elo (Code): 1557 **Q: When was Claude Opus 4.7 released?** A: Claude Opus 4.7 was released by Anthropic on Thursday, Apr 16 2026. **Q: Who made Claude Opus 4.7?** A: Claude Opus 4.7 was built by Anthropic. AI safety company building the Claude family of models. Founded in 2021 by former OpenAI researchers. **Q: What benchmark scores did Claude Opus 4.7 get?** A: Claude Opus 4.7 reports 27 tracked benchmark scores — BullshitBench v2: 83%; SWE-Bench Pro: 64.3%; SWE-Bench Verified: 87.6%; SWE-Bench Multilingual: 80.5%; CursorBench v3.1: 64.8%; Next.js Evals: 75%; Terminal-Bench 2.1: 66.1%; Terminal-Bench 2.0: 69.4%; MCP Atlas: 79.1%; BrowseComp: 79.3%; CyberGym: 73.1%; Humanity's Last Exam (no tools): 46.9%; Humanity's Last Exam (with tools): 54.7%; ARC-AGI-2: 75.8%; FrontierMath (Tier 1–3): 43.8%; FrontierMath (Tier 4): 22.9%; GPQA Diamond: 94.2%; OSWorld-Verified: 78%; Finance Agent v2: 51.5%; GDPval-AA: 1753; GDPval (win/tie rate): 80.3%; CharXiv Reasoning: 82.1%; MMMU-Pro: 75.2%; Blueprint-Bench 2: 24.5%; MRCR v2 (8-needle) (128k average): 59.3%; Arena Elo (Text): 1503; Arena Elo (Code): 1557. Scores are the figures published at release by Anthropic. **Q: What is the context window of Claude Opus 4.7?** A: Claude Opus 4.7 has a context window of 1M. That is the maximum amount of input plus output the model can hold in a single request. **Q: Is Claude Opus 4.7 open source?** A: No. Claude Opus 4.7 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Claude Opus 4.7?** A: Anthropic's previous tracked release was Claude Sonnet 4.6 on Feb 17 2026, 58 days earlier. It was followed by Claude Opus 4.8 on May 28 2026. ### Qwen3.6 - URL: https://aireleasetracker.com/model/qwen/qwen3.6 - Developer: Qwen - Release date: Thursday, Apr 16 2026 - Licensing: Open Weight - Parameters: 35B - Context window: 256k - Benchmarks published at release: - SWE-Bench Pro: 49.5% - SWE-Bench Verified: 73.4% - SWE-Bench Multilingual: 67.2% - Terminal-Bench 2.0: 51.5% - Humanity's Last Exam (no tools): 21.4% - GPQA Diamond: 86% - CharXiv Reasoning: 78% - MMMU-Pro: 75.3% - MMMU: 81.7% **Q: When was Qwen3.6 released?** A: Qwen3.6 was released by Qwen on Thursday, Apr 16 2026. **Q: Who made Qwen3.6?** A: Qwen3.6 was built by Qwen. Alibaba's AI lab, building the Qwen family. The most prolific publisher of open-weight models of any major lab, alongside a proprietary Max and Plus tier sold through Alibaba Cloud. **Q: What benchmark scores did Qwen3.6 get?** A: Qwen3.6 reports 9 tracked benchmark scores — SWE-Bench Pro: 49.5%; SWE-Bench Verified: 73.4%; SWE-Bench Multilingual: 67.2%; Terminal-Bench 2.0: 51.5%; Humanity's Last Exam (no tools): 21.4%; GPQA Diamond: 86%; CharXiv Reasoning: 78%; MMMU-Pro: 75.3%; MMMU: 81.7%. Scores are the figures published at release by Qwen. **Q: What is the context window of Qwen3.6?** A: Qwen3.6 has a context window of 256k. That is the maximum amount of input plus output the model can hold in a single request. **Q: How many parameters does Qwen3.6 have?** A: Qwen3.6 is reported at 35B parameters. **Q: Is Qwen3.6 open source?** A: Partly. Qwen3.6 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Qwen3.6?** A: Qwen's previous tracked release was Qwen3.6-Plus on Apr 2 2026, 14 days earlier. It was followed by Qwen3.7-Max on May 20 2026. ### Muse Spark - URL: https://aireleasetracker.com/model/meta/muse-spark - Developer: Meta - Release date: Wednesday, Apr 8 2026 - Licensing: Proprietary - Benchmarks published at release: - Gray Swan IPI (k = 1): 2.9% - Gray Swan IPI (k = 10): 14.3% - Gray Swan IPI (k = 15): 16.5% - SWE-Bench Pro: 55% - SWE-Bench Verified: 77.4% - DeepSWE 1.1: 10% - Terminal-Bench 2.1: 67.3% - MCP Atlas: 82.2% - JobBench: 17% - Toolathlon-Verified: 49.4% - Humanity's Last Exam (with tools): 50.4% - ARC-AGI-2: 42.5% - GPQA Diamond: 89.5% - OSWorld-Verified: 53.3% - CharXiv Reasoning: 88.9% - BabyVision: 39.9% - MMMU: 80.4% - Arena Elo (Text): 1488 Muse Spark, released April 8, 2026, was Meta's reset: a new model family retiring the Llama name a year after Llama 4, and — a sharp break with Meta's open-weights tradition — released as a proprietary model rather than a downloadable one. It debuted at 89.5% on GPQA Diamond, 80.4% on MMMU, and 77.4% on SWE-Bench Verified. Its standout results were agentic: 82.2% on MCP Atlas for tool orchestration and 88.9% on CharXiv Reasoning, the best chart-understanding score of any model at release. Muse Spark 1.1 followed on July 9, 2026 with large gains across agentic benchmarks, including 20.0% on Harvey Legal Agent — the top published legal-work score on this tracker, and the coding-focused Muse Spark 1.2 arrived a month later. **Q: When was Muse Spark released?** A: Muse Spark was released by Meta on Wednesday, Apr 8 2026. **Q: Who made Muse Spark?** A: Muse Spark was built by Meta. Develops the open-weight Llama series of models. Committed to open-source AI research. **Q: What benchmark scores did Muse Spark get?** A: Muse Spark reports 18 tracked benchmark scores — Gray Swan IPI (k = 1): 2.9%; Gray Swan IPI (k = 10): 14.3%; Gray Swan IPI (k = 15): 16.5%; SWE-Bench Pro: 55%; SWE-Bench Verified: 77.4%; DeepSWE 1.1: 10%; Terminal-Bench 2.1: 67.3%; MCP Atlas: 82.2%; JobBench: 17%; Toolathlon-Verified: 49.4%; Humanity's Last Exam (with tools): 50.4%; ARC-AGI-2: 42.5%; GPQA Diamond: 89.5%; OSWorld-Verified: 53.3%; CharXiv Reasoning: 88.9%; BabyVision: 39.9%; MMMU: 80.4%; Arena Elo (Text): 1488. Scores are the figures published at release by Meta. It holds the best score among all models tracked here on CharXiv Reasoning. **Q: Is Muse Spark open source?** A: No. Muse Spark is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Muse Spark?** A: Meta's previous tracked release was LLaMA 4 Maverick on Apr 5 2025, 368 days earlier. It was followed by Muse Spark 1.1 on Jul 9 2026. ### GLM-5.1 - URL: https://aireleasetracker.com/model/zai/glm-5.1 - Developer: Z.ai - Release date: Tuesday, Apr 7 2026 - Licensing: Open Weight - Parameters: 744B - Context window: 200k - Benchmarks published at release: - BullshitBench v2: 22% - Next.js Evals: 75% - Terminal-Bench 2.0: 63.5% - BrowseComp: 68% - Humanity's Last Exam (with tools): 52.3% - GPQA Diamond: 86.2% - Arena Elo (Text): 1468 - Arena Elo (Code): 1518 **Q: When was GLM-5.1 released?** A: GLM-5.1 was released by Z.ai on Tuesday, Apr 7 2026. **Q: Who made GLM-5.1?** A: GLM-5.1 was built by Z.ai. Chinese AI lab spun out of Tsinghua University (formerly Zhipu AI), building the open-weight GLM family. Rebranded internationally as Z.ai in 2025. **Q: What benchmark scores did GLM-5.1 get?** A: GLM-5.1 reports 8 tracked benchmark scores — BullshitBench v2: 22%; Next.js Evals: 75%; Terminal-Bench 2.0: 63.5%; BrowseComp: 68%; Humanity's Last Exam (with tools): 52.3%; GPQA Diamond: 86.2%; Arena Elo (Text): 1468; Arena Elo (Code): 1518. Scores are the figures published at release by Z.ai. **Q: What is the context window of GLM-5.1?** A: GLM-5.1 has a context window of 200k. That is the maximum amount of input plus output the model can hold in a single request. **Q: How many parameters does GLM-5.1 have?** A: GLM-5.1 is reported at 744B parameters. **Q: Is GLM-5.1 open source?** A: Partly. GLM-5.1 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after GLM-5.1?** A: Z.ai's previous tracked release was GLM-5 on Feb 12 2026, 54 days earlier. It was followed by GLM-5.2 on Jun 16 2026. ### Gemma 4 - URL: https://aireleasetracker.com/model/google/gemma-4 - Developer: Google - Release date: Thursday, Apr 2 2026 - Licensing: Open Weight - Benchmarks: none published in this dataset **Q: When was Gemma 4 released?** A: Gemma 4 was released by Google on Thursday, Apr 2 2026. **Q: Who made Gemma 4?** A: Gemma 4 was built by Google. Builds the Gemini family of models through Google DeepMind. Integrates AI across Google products. **Q: Is Gemma 4 open source?** A: Partly. Gemma 4 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Gemma 4?** A: Google's previous tracked release was Gemini 3.1 Flash-Lite on Mar 3 2026, 30 days earlier. It was followed by Gemini 3.5 Flash on May 19 2026. ### Qwen3.6-Plus - URL: https://aireleasetracker.com/model/qwen/qwen3.6-plus - Developer: Qwen - Release date: Thursday, Apr 2 2026 - Licensing: Proprietary - Benchmarks published at release: - BullshitBench v2: 72% - Arena Elo (Code): 1458 **Q: When was Qwen3.6-Plus released?** A: Qwen3.6-Plus was released by Qwen on Thursday, Apr 2 2026. **Q: Who made Qwen3.6-Plus?** A: Qwen3.6-Plus was built by Qwen. Alibaba's AI lab, building the Qwen family. The most prolific publisher of open-weight models of any major lab, alongside a proprietary Max and Plus tier sold through Alibaba Cloud. **Q: What benchmark scores did Qwen3.6-Plus get?** A: Qwen3.6-Plus reports 2 tracked benchmark scores — BullshitBench v2: 72%; Arena Elo (Code): 1458. Scores are the figures published at release by Qwen. **Q: Is Qwen3.6-Plus open source?** A: No. Qwen3.6-Plus is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Qwen3.6-Plus?** A: Qwen's previous tracked release was Qwen3.5-Plus on Feb 16 2026, 45 days earlier. It was followed by Qwen3.6 on Apr 16 2026. ### Composer 2 - URL: https://aireleasetracker.com/model/cursor/composer-2 - Developer: Cursor - Release date: Thursday, Mar 19 2026 - Licensing: Proprietary - Benchmarks published at release: - SWE-Bench Multilingual: 73.7% - CursorBench v3.1: 52.2% - Next.js Evals: 75% - Terminal-Bench 2.0: 61.7% **Q: When was Composer 2 released?** A: Composer 2 was released by Cursor on Thursday, Mar 19 2026. **Q: Who made Composer 2?** A: Composer 2 was built by Cursor. **Q: What benchmark scores did Composer 2 get?** A: Composer 2 reports 4 tracked benchmark scores — SWE-Bench Multilingual: 73.7%; CursorBench v3.1: 52.2%; Next.js Evals: 75%; Terminal-Bench 2.0: 61.7%. Scores are the figures published at release by Cursor. **Q: Is Composer 2 open source?** A: No. Composer 2 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Composer 2?** A: Cursor's previous tracked release was Composer 1.5 on Feb 9 2026, 38 days earlier. It was followed by Composer 2.5 on May 18 2026. ### GPT-5.4 mini - URL: https://aireleasetracker.com/model/openai/gpt-5.4-mini - Developer: OpenAI - Release date: Tuesday, Mar 17 2026 - Licensing: Proprietary - Context window: 400k - Benchmarks published at release: - BullshitBench v2: 32% - Supabase Evals (with skills): 81.8% - Supabase Evals (no skills): 63.6% - BU Bench: 36% - Arena Elo (Code): 1398 **Q: When was GPT-5.4 mini released?** A: GPT-5.4 mini was released by OpenAI on Tuesday, Mar 17 2026. **Q: Who made GPT-5.4 mini?** A: GPT-5.4 mini was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: What benchmark scores did GPT-5.4 mini get?** A: GPT-5.4 mini reports 5 tracked benchmark scores — BullshitBench v2: 32%; Supabase Evals (with skills): 81.8%; Supabase Evals (no skills): 63.6%; BU Bench: 36%; Arena Elo (Code): 1398. Scores are the figures published at release by OpenAI. **Q: What is the context window of GPT-5.4 mini?** A: GPT-5.4 mini has a context window of 400k. That is the maximum amount of input plus output the model can hold in a single request. **Q: Is GPT-5.4 mini open source?** A: No. GPT-5.4 mini is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after GPT-5.4 mini?** A: OpenAI's previous tracked release was GPT-5.4-Pro on Mar 5 2026, 12 days earlier. It was followed by GPT-5.4 nano on Mar 17 2026. ### GPT-5.4 nano - URL: https://aireleasetracker.com/model/openai/gpt-5.4-nano - Developer: OpenAI - Release date: Tuesday, Mar 17 2026 - Licensing: Proprietary - Benchmarks published at release: - BullshitBench v2: 14% **Q: When was GPT-5.4 nano released?** A: GPT-5.4 nano was released by OpenAI on Tuesday, Mar 17 2026. **Q: Who made GPT-5.4 nano?** A: GPT-5.4 nano was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: What benchmark scores did GPT-5.4 nano get?** A: GPT-5.4 nano reports 1 tracked benchmark score — BullshitBench v2: 14%. Scores are the figures published at release by OpenAI. **Q: Is GPT-5.4 nano open source?** A: No. GPT-5.4 nano is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after GPT-5.4 nano?** A: OpenAI's previous tracked release was GPT-5.4 mini on Mar 17 2026. It was followed by GPT-5.5 on Apr 23 2026. ### Mistral Small 4 - URL: https://aireleasetracker.com/model/mistral/mistral-small-4 - Developer: Mistral - Release date: Monday, Mar 16 2026 - Licensing: Proprietary - Benchmarks: none published in this dataset **Q: When was Mistral Small 4 released?** A: Mistral Small 4 was released by Mistral on Monday, Mar 16 2026. **Q: Who made Mistral Small 4?** A: Mistral Small 4 was built by Mistral. French AI company building open and commercial models. Founded in 2023 by former Meta and DeepMind researchers. **Q: Is Mistral Small 4 open source?** A: No. Mistral Small 4 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Mistral Small 4?** A: Mistral's previous tracked release was Devstral Small 2 (24B) on Dec 9 2025, 97 days earlier. It was followed by Mistral Medium 3.5 on Apr 29 2026. ### GPT-5.4 - URL: https://aireleasetracker.com/model/openai/gpt-5.4 - Developer: OpenAI - Release date: Thursday, Mar 5 2026 - Licensing: Proprietary - Benchmarks published at release: - BullshitBench v2: 48% - Next.js Evals: 83% - Terminal-Bench 2.0: 75.1% - Expert-SWE (Internal): 68.5% - Toolathlon: 54.6% - BrowseComp: 82.7% - CyberGym: 79% - FrontierMath (Tier 1–3): 47.6% - FrontierMath (Tier 4): 27.1% - GPQA Diamond: 92.8% - OSWorld-Verified: 75% - GDPval (win/tie rate): 83% - Arena Elo (Text): 1476 - Arena Elo (Code): 1462 **Q: When was GPT-5.4 released?** A: GPT-5.4 was released by OpenAI on Thursday, Mar 5 2026. **Q: Who made GPT-5.4?** A: GPT-5.4 was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: What benchmark scores did GPT-5.4 get?** A: GPT-5.4 reports 14 tracked benchmark scores — BullshitBench v2: 48%; Next.js Evals: 83%; Terminal-Bench 2.0: 75.1%; Expert-SWE (Internal): 68.5%; Toolathlon: 54.6%; BrowseComp: 82.7%; CyberGym: 79%; FrontierMath (Tier 1–3): 47.6%; FrontierMath (Tier 4): 27.1%; GPQA Diamond: 92.8%; OSWorld-Verified: 75%; GDPval (win/tie rate): 83%; Arena Elo (Text): 1476; Arena Elo (Code): 1462. Scores are the figures published at release by OpenAI. **Q: Is GPT-5.4 open source?** A: No. GPT-5.4 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after GPT-5.4?** A: OpenAI's previous tracked release was GPT-5.3-Codex-Spark on Feb 12 2026, 21 days earlier. It was followed by GPT-5.4-Pro on Mar 5 2026. ### GPT-5.4-Pro - URL: https://aireleasetracker.com/model/openai/gpt-5.4-pro - Developer: OpenAI - Release date: Thursday, Mar 5 2026 - Licensing: Proprietary - Benchmarks published at release: - BrowseComp: 89.3% - Humanity's Last Exam (with tools): 58.7% - ARC-AGI-2: 83.3% - FrontierMath (Tier 1–3): 50% - FrontierMath (Tier 4): 38% - GPQA Diamond: 94.4% - GDPval (win/tie rate): 82% **Q: When was GPT-5.4-Pro released?** A: GPT-5.4-Pro was released by OpenAI on Thursday, Mar 5 2026. **Q: Who made GPT-5.4-Pro?** A: GPT-5.4-Pro was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: What benchmark scores did GPT-5.4-Pro get?** A: GPT-5.4-Pro reports 7 tracked benchmark scores — BrowseComp: 89.3%; Humanity's Last Exam (with tools): 58.7%; ARC-AGI-2: 83.3%; FrontierMath (Tier 1–3): 50%; FrontierMath (Tier 4): 38%; GPQA Diamond: 94.4%; GDPval (win/tie rate): 82%. Scores are the figures published at release by OpenAI. It holds the best score among all models tracked here on GPQA Diamond. **Q: Is GPT-5.4-Pro open source?** A: No. GPT-5.4-Pro is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after GPT-5.4-Pro?** A: OpenAI's previous tracked release was GPT-5.4 on Mar 5 2026. It was followed by GPT-5.4 mini on Mar 17 2026. ### Gemini 3.1 Flash-Lite - URL: https://aireleasetracker.com/model/google/gemini-3.1-flash-lite - Developer: Google - Release date: Tuesday, Mar 3 2026 - Licensing: Proprietary - Benchmarks published at release: - BullshitBench v2: 11% - Arena Elo (Code): 1253 **Q: When was Gemini 3.1 Flash-Lite released?** A: Gemini 3.1 Flash-Lite was released by Google on Tuesday, Mar 3 2026. **Q: Who made Gemini 3.1 Flash-Lite?** A: Gemini 3.1 Flash-Lite was built by Google. Builds the Gemini family of models through Google DeepMind. Integrates AI across Google products. **Q: What benchmark scores did Gemini 3.1 Flash-Lite get?** A: Gemini 3.1 Flash-Lite reports 2 tracked benchmark scores — BullshitBench v2: 11%; Arena Elo (Code): 1253. Scores are the figures published at release by Google. **Q: Is Gemini 3.1 Flash-Lite open source?** A: No. Gemini 3.1 Flash-Lite is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Gemini 3.1 Flash-Lite?** A: Google's previous tracked release was Gemini 3.1 Pro on Feb 19 2026, 12 days earlier. It was followed by Gemma 4 on Apr 2 2026. ### Gemini 3.1 Pro - URL: https://aireleasetracker.com/model/google/gemini-3.1-pro - Developer: Google - Release date: Thursday, Feb 19 2026 - Licensing: Proprietary - Benchmarks published at release: - BullshitBench v2: 37% - Gray Swan IPI (k = 1): 14.2% - Gray Swan IPI (k = 10): 45.7% - Gray Swan IPI (k = 15): 49.2% - SWE-Bench Pro: 54.2% - SWE-Bench Verified: 80.6% - DeepSWE 1.1: 12% - MLE-Bench: 42.6% - Next.js Evals: 75% - Terminal-Bench 2.1: 70.3% - Terminal-Bench 2.0: 68.5% - MCP Atlas: 78.2% - Toolathlon: 48.8% - BrowseComp: 85.9% - Humanity's Last Exam (no tools): 44.4% - Humanity's Last Exam (with tools): 51.4% - ARC-AGI-2: 77.1% - FrontierMath (Tier 1–3): 36.9% - FrontierMath (Tier 4): 16.7% - GPQA Diamond: 94.3% - OSWorld-Verified: 76.2% - Finance Agent v2: 43% - GDPval-AA: 1314 - GDPval-AA v2: 965 - GDPval (win/tie rate): 67.3% - CharXiv Reasoning: 83.3% - MMMU-Pro: 80.5% - Blueprint-Bench 2: 26.5% - MRCR v2 (8-needle) (128k average): 84.9% - MRCR v2 (8-needle) (1M pointwise): 26.3% - Arena Elo (Text): 1485 - Arena Elo (Code): 1445 Gemini 3.1 Pro, released February 19, 2026, sharpened the 3-series flagship into a reasoning specialist: 94.3% on GPQA Diamond and 77.1% on ARC-AGI-2 were both the top published scores at release, alongside 85.9% on BrowseComp for agentic web research and 84.9% on MRCR v2 128K long-context recall. It landed in the same fortnight as Anthropic's Claude Sonnet 4.6 and xAI's Grok 4.20 Beta, in one of 2026's densest release windows. The 3.1 generation filled out below it with Gemini 3.1 Flash-Lite in early March, before Google's attention shifted to the 3.5 Flash line in May 2026. **Q: When was Gemini 3.1 Pro released?** A: Gemini 3.1 Pro was released by Google on Thursday, Feb 19 2026. **Q: Who made Gemini 3.1 Pro?** A: Gemini 3.1 Pro was built by Google. Builds the Gemini family of models through Google DeepMind. Integrates AI across Google products. **Q: What benchmark scores did Gemini 3.1 Pro get?** A: Gemini 3.1 Pro reports 32 tracked benchmark scores — BullshitBench v2: 37%; Gray Swan IPI (k = 1): 14.2%; Gray Swan IPI (k = 10): 45.7%; Gray Swan IPI (k = 15): 49.2%; SWE-Bench Pro: 54.2%; SWE-Bench Verified: 80.6%; DeepSWE 1.1: 12%; MLE-Bench: 42.6%; Next.js Evals: 75%; Terminal-Bench 2.1: 70.3%; Terminal-Bench 2.0: 68.5%; MCP Atlas: 78.2%; Toolathlon: 48.8%; BrowseComp: 85.9%; Humanity's Last Exam (no tools): 44.4%; Humanity's Last Exam (with tools): 51.4%; ARC-AGI-2: 77.1%; FrontierMath (Tier 1–3): 36.9%; FrontierMath (Tier 4): 16.7%; GPQA Diamond: 94.3%; OSWorld-Verified: 76.2%; Finance Agent v2: 43%; GDPval-AA: 1314; GDPval-AA v2: 965; GDPval (win/tie rate): 67.3%; CharXiv Reasoning: 83.3%; MMMU-Pro: 80.5%; Blueprint-Bench 2: 26.5%; MRCR v2 (8-needle) (128k average): 84.9%; MRCR v2 (8-needle) (1M pointwise): 26.3%; Arena Elo (Text): 1485; Arena Elo (Code): 1445. Scores are the figures published at release by Google. **Q: Is Gemini 3.1 Pro open source?** A: No. Gemini 3.1 Pro is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Gemini 3.1 Pro?** A: Google's previous tracked release was Gemini 3.0 Flash on Dec 17 2025, 64 days earlier. It was followed by Gemini 3.1 Flash-Lite on Mar 3 2026. ### Grok 4.20 Beta - URL: https://aireleasetracker.com/model/xai/grok-4.20-beta - Developer: SpaceXAI - Release date: Tuesday, Feb 17 2026 - Licensing: Proprietary - Benchmarks published at release: - BullshitBench v2: 56% - ARC-AGI-2: 53.3% - Arena Elo (Text): 1475 **Q: When was Grok 4.20 Beta released?** A: Grok 4.20 Beta was released by SpaceXAI on Tuesday, Feb 17 2026. **Q: Who made Grok 4.20 Beta?** A: Grok 4.20 Beta was built by SpaceXAI. Elon Musk's AI company building the Grok series of models. Founded in 2023 as xAI, now part of SpaceX. **Q: What benchmark scores did Grok 4.20 Beta get?** A: Grok 4.20 Beta reports 3 tracked benchmark scores — BullshitBench v2: 56%; ARC-AGI-2: 53.3%; Arena Elo (Text): 1475. Scores are the figures published at release by SpaceXAI. **Q: Is Grok 4.20 Beta open source?** A: No. Grok 4.20 Beta is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Grok 4.20 Beta?** A: SpaceXAI's previous tracked release was Grok 4.1 Fast on Nov 19 2025, 90 days earlier. It was followed by Grok 4.3 Beta on Apr 17 2026. ### Claude Sonnet 4.6 - URL: https://aireleasetracker.com/model/anthropic/claude-sonnet-4.6 - Developer: Anthropic - Release date: Tuesday, Feb 17 2026 - Licensing: Proprietary - Benchmarks published at release: - BullshitBench v2: 91% - SWE-Bench Verified: 79.6% - CursorBench v3.1: 49% - DeepSWE 1.1: 30% - Next.js Evals: 58% - MCP Atlas: 69.5% - BU Bench: 62% - Humanity's Last Exam (no tools): 33.2% - ARC-AGI-2: 58.3% - GPQA Diamond: 89.9% - OSWorld-Verified: 72.5% - Finance Agent v2: 51% - GDPval-AA: 1676 - CharXiv Reasoning: 72.4% - MMMU-Pro: 74.5% - Blueprint-Bench 2: 6.7% - MRCR v2 (8-needle) (128k average): 84.9% - Arena Elo (Text): 1472 - Arena Elo (Code): 1521 **Q: When was Claude Sonnet 4.6 released?** A: Claude Sonnet 4.6 was released by Anthropic on Tuesday, Feb 17 2026. **Q: Who made Claude Sonnet 4.6?** A: Claude Sonnet 4.6 was built by Anthropic. AI safety company building the Claude family of models. Founded in 2021 by former OpenAI researchers. **Q: What benchmark scores did Claude Sonnet 4.6 get?** A: Claude Sonnet 4.6 reports 19 tracked benchmark scores — BullshitBench v2: 91%; SWE-Bench Verified: 79.6%; CursorBench v3.1: 49%; DeepSWE 1.1: 30%; Next.js Evals: 58%; MCP Atlas: 69.5%; BU Bench: 62%; Humanity's Last Exam (no tools): 33.2%; ARC-AGI-2: 58.3%; GPQA Diamond: 89.9%; OSWorld-Verified: 72.5%; Finance Agent v2: 51%; GDPval-AA: 1676; CharXiv Reasoning: 72.4%; MMMU-Pro: 74.5%; Blueprint-Bench 2: 6.7%; MRCR v2 (8-needle) (128k average): 84.9%; Arena Elo (Text): 1472; Arena Elo (Code): 1521. Scores are the figures published at release by Anthropic. **Q: Is Claude Sonnet 4.6 open source?** A: No. Claude Sonnet 4.6 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Claude Sonnet 4.6?** A: Anthropic's previous tracked release was Claude Opus 4.6 on Feb 5 2026, 12 days earlier. It was followed by Claude Opus 4.7 on Apr 16 2026. ### Qwen3.5 - URL: https://aireleasetracker.com/model/qwen/qwen3.5 - Developer: Qwen - Release date: Monday, Feb 16 2026 - Licensing: Open Weight - Parameters: 397B - Context window: 1M - Benchmarks published at release: - SWE-Bench Verified: 76.4% - SWE-Bench Multilingual: 69.3% - Terminal-Bench 2.0: 52.5% - BrowseComp: 69% - Humanity's Last Exam (no tools): 28.7% - GPQA Diamond: 88.4% - OSWorld-Verified: 62.2% - CharXiv Reasoning: 80.8% - MMMU-Pro: 79% - MMMU: 85% Qwen3.5, released February 16, 2026, was a 397B-parameter mixture-of-experts model with 17B active, published under Apache 2.0 with a 1M-token context window. Alibaba framed it around agent work rather than chat: the launch claim was that it could operate desktop and mobile applications directly, which put it in the same competitive frame as the computer-use models the US labs had been shipping. On Alibaba's own harness it posted 76.4% on SWE-Bench Verified, 88.4% on GPQA Diamond and 62.2% on OSWorld-Verified. The last of those fit the agent pitch — GPT-5.2 managed 38.2% on the same table — even as Qwen3.5 trailed GPT-5.2 and Claude Opus 4.5 by roughly four points on SWE-Bench Verified and sat at 52.5% on Terminal-Bench 2. It launched the same day as the proprietary Qwen3.5-Plus, making explicit the two-tier structure Qwen3-Max had introduced five months earlier. Smaller open members of the family followed quickly — 122B-A10B, 35B-A3B and 27B on February 24, then 9B, 4B, 2B and 0.8B variants on March 2, 2026 — repeating the size-ladder pattern that had made Qwen2.5 and Qwen3 so widely adopted. **Q: When was Qwen3.5 released?** A: Qwen3.5 was released by Qwen on Monday, Feb 16 2026. **Q: Who made Qwen3.5?** A: Qwen3.5 was built by Qwen. Alibaba's AI lab, building the Qwen family. The most prolific publisher of open-weight models of any major lab, alongside a proprietary Max and Plus tier sold through Alibaba Cloud. **Q: What benchmark scores did Qwen3.5 get?** A: Qwen3.5 reports 10 tracked benchmark scores — SWE-Bench Verified: 76.4%; SWE-Bench Multilingual: 69.3%; Terminal-Bench 2.0: 52.5%; BrowseComp: 69%; Humanity's Last Exam (no tools): 28.7%; GPQA Diamond: 88.4%; OSWorld-Verified: 62.2%; CharXiv Reasoning: 80.8%; MMMU-Pro: 79%; MMMU: 85%. Scores are the figures published at release by Qwen. It holds the best score among all models tracked here on MMMU. **Q: What is the context window of Qwen3.5?** A: Qwen3.5 has a context window of 1M. That is the maximum amount of input plus output the model can hold in a single request. **Q: How many parameters does Qwen3.5 have?** A: Qwen3.5 is reported at 397B parameters. **Q: Is Qwen3.5 open source?** A: Partly. Qwen3.5 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Qwen3.5?** A: Qwen's previous tracked release was Qwen3-Coder-Next on Feb 3 2026, 13 days earlier. It was followed by Qwen3.5-Plus on Feb 16 2026. ### Qwen3.5-Plus - URL: https://aireleasetracker.com/model/qwen/qwen3.5-plus - Developer: Qwen - Release date: Monday, Feb 16 2026 - Licensing: Proprietary - Context window: 1M - Benchmarks: none published in this dataset **Q: When was Qwen3.5-Plus released?** A: Qwen3.5-Plus was released by Qwen on Monday, Feb 16 2026. **Q: Who made Qwen3.5-Plus?** A: Qwen3.5-Plus was built by Qwen. Alibaba's AI lab, building the Qwen family. The most prolific publisher of open-weight models of any major lab, alongside a proprietary Max and Plus tier sold through Alibaba Cloud. **Q: What is the context window of Qwen3.5-Plus?** A: Qwen3.5-Plus has a context window of 1M. That is the maximum amount of input plus output the model can hold in a single request. **Q: Is Qwen3.5-Plus open source?** A: No. Qwen3.5-Plus is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Qwen3.5-Plus?** A: Qwen's previous tracked release was Qwen3.5 on Feb 16 2026. It was followed by Qwen3.6-Plus on Apr 2 2026. ### GPT-5.3-Codex-Spark - URL: https://aireleasetracker.com/model/openai/gpt-5.3-codex-spark - Developer: OpenAI - Release date: Thursday, Feb 12 2026 - Licensing: Proprietary - Context window: 128k - Benchmarks: none published in this dataset **Q: When was GPT-5.3-Codex-Spark released?** A: GPT-5.3-Codex-Spark was released by OpenAI on Thursday, Feb 12 2026. **Q: Who made GPT-5.3-Codex-Spark?** A: GPT-5.3-Codex-Spark was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: What is the context window of GPT-5.3-Codex-Spark?** A: GPT-5.3-Codex-Spark has a context window of 128k. That is the maximum amount of input plus output the model can hold in a single request. **Q: Is GPT-5.3-Codex-Spark open source?** A: No. GPT-5.3-Codex-Spark is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after GPT-5.3-Codex-Spark?** A: OpenAI's previous tracked release was GPT-5.3-Codex on Feb 5 2026, 7 days earlier. It was followed by GPT-5.4 on Mar 5 2026. ### GLM-5 - URL: https://aireleasetracker.com/model/zai/glm-5 - Developer: Z.ai - Release date: Thursday, Feb 12 2026 - Licensing: Open Weight - Parameters: 744B - Benchmarks published at release: - BullshitBench v2: 28% - SWE-Bench Verified: 77.8% - SWE-Bench Multilingual: 73.3% - Terminal-Bench 2.0: 56.2% - BrowseComp: 75.9% - Humanity's Last Exam (with tools): 50.4% - GPQA Diamond: 86% - Arena Elo (Code): 1435 GLM-5, released February 12, 2026, doubled Z.ai's flagship to 744B parameters while keeping the weights open. It scored 77.8% on SWE-Bench Verified, 86.0% on GPQA, and 75.9% on BrowseComp — at release the strongest agentic-research score of any open-weight model — with 50.4% on Humanity's Last Exam with tools. Landing a week after Claude Opus 4.6 and GPT-5.3-Codex, GLM-5 kept open weights within striking distance of the closed frontier through early 2026. GLM-5.1 followed in April with a 200K context window, and GLM-5.2 pushed the family to a 1M-token context in June 2026. **Q: When was GLM-5 released?** A: GLM-5 was released by Z.ai on Thursday, Feb 12 2026. **Q: Who made GLM-5?** A: GLM-5 was built by Z.ai. Chinese AI lab spun out of Tsinghua University (formerly Zhipu AI), building the open-weight GLM family. Rebranded internationally as Z.ai in 2025. **Q: What benchmark scores did GLM-5 get?** A: GLM-5 reports 8 tracked benchmark scores — BullshitBench v2: 28%; SWE-Bench Verified: 77.8%; SWE-Bench Multilingual: 73.3%; Terminal-Bench 2.0: 56.2%; BrowseComp: 75.9%; Humanity's Last Exam (with tools): 50.4%; GPQA Diamond: 86%; Arena Elo (Code): 1435. Scores are the figures published at release by Z.ai. **Q: How many parameters does GLM-5 have?** A: GLM-5 is reported at 744B parameters. **Q: Is GLM-5 open source?** A: Partly. GLM-5 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after GLM-5?** A: Z.ai's previous tracked release was GLM-4.7 on Dec 22 2025, 52 days earlier. It was followed by GLM-5.1 on Apr 7 2026. ### Composer 1.5 - URL: https://aireleasetracker.com/model/cursor/composer-1.5 - Developer: Cursor - Release date: Monday, Feb 9 2026 - Licensing: Proprietary - Benchmarks published at release: - Next.js Evals: 67% **Q: When was Composer 1.5 released?** A: Composer 1.5 was released by Cursor on Monday, Feb 9 2026. **Q: Who made Composer 1.5?** A: Composer 1.5 was built by Cursor. **Q: What benchmark scores did Composer 1.5 get?** A: Composer 1.5 reports 1 tracked benchmark score — Next.js Evals: 67%. Scores are the figures published at release by Cursor. **Q: Is Composer 1.5 open source?** A: No. Composer 1.5 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Composer 1.5?** A: Cursor's previous tracked release was Composer on Oct 29 2025, 103 days earlier. It was followed by Composer 2 on Mar 19 2026. ### Claude Opus 4.6 - URL: https://aireleasetracker.com/model/anthropic/claude-opus-4.6 - Developer: Anthropic - Release date: Thursday, Feb 5 2026 - Licensing: Proprietary - Benchmarks published at release: - BullshitBench v2: 87% - SWE-Bench Verified: 80.8% - Next.js Evals: 75% - BrowseComp: 83.7% - Humanity's Last Exam (with tools): 53% - GPQA Diamond: 91.3% - Arena Elo (Text): 1504 - Arena Elo (Code): 1543 **Q: When was Claude Opus 4.6 released?** A: Claude Opus 4.6 was released by Anthropic on Thursday, Feb 5 2026. **Q: Who made Claude Opus 4.6?** A: Claude Opus 4.6 was built by Anthropic. AI safety company building the Claude family of models. Founded in 2021 by former OpenAI researchers. **Q: What benchmark scores did Claude Opus 4.6 get?** A: Claude Opus 4.6 reports 8 tracked benchmark scores — BullshitBench v2: 87%; SWE-Bench Verified: 80.8%; Next.js Evals: 75%; BrowseComp: 83.7%; Humanity's Last Exam (with tools): 53%; GPQA Diamond: 91.3%; Arena Elo (Text): 1504; Arena Elo (Code): 1543. Scores are the figures published at release by Anthropic. **Q: Is Claude Opus 4.6 open source?** A: No. Claude Opus 4.6 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Claude Opus 4.6?** A: Anthropic's previous tracked release was Claude Opus 4.5 on Nov 24 2025, 73 days earlier. It was followed by Claude Sonnet 4.6 on Feb 17 2026. ### GPT-5.3-Codex - URL: https://aireleasetracker.com/model/openai/gpt-5.3-codex - Developer: OpenAI - Release date: Thursday, Feb 5 2026 - Licensing: Proprietary - Benchmarks published at release: - BullshitBench v2: 24% - Next.js Evals: 83% - Arena Elo (Code): 1406 **Q: When was GPT-5.3-Codex released?** A: GPT-5.3-Codex was released by OpenAI on Thursday, Feb 5 2026. **Q: Who made GPT-5.3-Codex?** A: GPT-5.3-Codex was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: What benchmark scores did GPT-5.3-Codex get?** A: GPT-5.3-Codex reports 3 tracked benchmark scores — BullshitBench v2: 24%; Next.js Evals: 83%; Arena Elo (Code): 1406. Scores are the figures published at release by OpenAI. **Q: Is GPT-5.3-Codex open source?** A: No. GPT-5.3-Codex is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after GPT-5.3-Codex?** A: OpenAI's previous tracked release was GPT-5.2 on Dec 11 2025, 56 days earlier. It was followed by GPT-5.3-Codex-Spark on Feb 12 2026. ### Qwen3-Coder-Next - URL: https://aireleasetracker.com/model/qwen/qwen3-coder-next - Developer: Qwen - Release date: Tuesday, Feb 3 2026 - Licensing: Open Weight - Parameters: 80B - Context window: 256k - Benchmarks published at release: - SWE-Bench Pro: 44.3% - SWE-Bench Verified: 70.6% - SWE-Bench Multilingual: 62.8% - Terminal-Bench 2.0: 36.2% **Q: When was Qwen3-Coder-Next released?** A: Qwen3-Coder-Next was released by Qwen on Tuesday, Feb 3 2026. **Q: Who made Qwen3-Coder-Next?** A: Qwen3-Coder-Next was built by Qwen. Alibaba's AI lab, building the Qwen family. The most prolific publisher of open-weight models of any major lab, alongside a proprietary Max and Plus tier sold through Alibaba Cloud. **Q: What benchmark scores did Qwen3-Coder-Next get?** A: Qwen3-Coder-Next reports 4 tracked benchmark scores — SWE-Bench Pro: 44.3%; SWE-Bench Verified: 70.6%; SWE-Bench Multilingual: 62.8%; Terminal-Bench 2.0: 36.2%. Scores are the figures published at release by Qwen. **Q: What is the context window of Qwen3-Coder-Next?** A: Qwen3-Coder-Next has a context window of 256k. That is the maximum amount of input plus output the model can hold in a single request. **Q: How many parameters does Qwen3-Coder-Next have?** A: Qwen3-Coder-Next is reported at 80B parameters. **Q: Is Qwen3-Coder-Next open source?** A: Partly. Qwen3-Coder-Next is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Qwen3-Coder-Next?** A: Qwen's previous tracked release was Qwen3-Max on Sep 24 2025, 132 days earlier. It was followed by Qwen3.5 on Feb 16 2026. ### Kimi K2.5 - URL: https://aireleasetracker.com/model/moonshot/kimi-k2.5 - Developer: Moonshot AI - Release date: Tuesday, Jan 27 2026 - Licensing: Open Weight - Parameters: 1T - Context window: 256k - Benchmarks published at release: - BullshitBench v2: 52% - SWE-Bench Verified: 76.8% - CursorBench v3.1: 31.9% - Next.js Evals: 21% - LiveCodeBench: 85% - GPQA Diamond: 87.6% - Arena Elo (Code): 1433 **Q: When was Kimi K2.5 released?** A: Kimi K2.5 was released by Moonshot AI on Tuesday, Jan 27 2026. **Q: Who made Kimi K2.5?** A: Kimi K2.5 was built by Moonshot AI. Chinese AI company building the Kimi family of models. Founded in 2023, known for long-context and efficient MoE architectures. **Q: What benchmark scores did Kimi K2.5 get?** A: Kimi K2.5 reports 7 tracked benchmark scores — BullshitBench v2: 52%; SWE-Bench Verified: 76.8%; CursorBench v3.1: 31.9%; Next.js Evals: 21%; LiveCodeBench: 85%; GPQA Diamond: 87.6%; Arena Elo (Code): 1433. Scores are the figures published at release by Moonshot AI. **Q: What is the context window of Kimi K2.5?** A: Kimi K2.5 has a context window of 256k. That is the maximum amount of input plus output the model can hold in a single request. **Q: How many parameters does Kimi K2.5 have?** A: Kimi K2.5 is reported at 1T parameters. **Q: Is Kimi K2.5 open source?** A: Partly. Kimi K2.5 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Kimi K2.5?** A: Moonshot AI's previous tracked release was Kimi K2 Thinking on Nov 6 2025, 82 days earlier. It was followed by Kimi K2.6 on Apr 21 2026. ### GLM-4.7 - URL: https://aireleasetracker.com/model/zai/glm-4.7 - Developer: Z.ai - Release date: Monday, Dec 22 2025 - Licensing: Open Weight - Context window: 128k - Benchmarks published at release: - SWE-Bench Verified: 73.8% - Terminal-Bench 2.0: 41% - BrowseComp: 52% - Humanity's Last Exam (no tools): 24.8% - Humanity's Last Exam (with tools): 42.8% - GPQA Diamond: 85.7% - Arena Elo (Code): 1433 **Q: When was GLM-4.7 released?** A: GLM-4.7 was released by Z.ai on Monday, Dec 22 2025. **Q: Who made GLM-4.7?** A: GLM-4.7 was built by Z.ai. Chinese AI lab spun out of Tsinghua University (formerly Zhipu AI), building the open-weight GLM family. Rebranded internationally as Z.ai in 2025. **Q: What benchmark scores did GLM-4.7 get?** A: GLM-4.7 reports 7 tracked benchmark scores — SWE-Bench Verified: 73.8%; Terminal-Bench 2.0: 41%; BrowseComp: 52%; Humanity's Last Exam (no tools): 24.8%; Humanity's Last Exam (with tools): 42.8%; GPQA Diamond: 85.7%; Arena Elo (Code): 1433. Scores are the figures published at release by Z.ai. **Q: What is the context window of GLM-4.7?** A: GLM-4.7 has a context window of 128k. That is the maximum amount of input plus output the model can hold in a single request. **Q: Is GLM-4.7 open source?** A: Partly. GLM-4.7 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after GLM-4.7?** A: Z.ai's previous tracked release was GLM-4.6 on Sep 30 2025, 83 days earlier. It was followed by GLM-5 on Feb 12 2026. ### Gemini 3.0 Flash - URL: https://aireleasetracker.com/model/google/gemini-3.0-flash - Developer: Google - Release date: Wednesday, Dec 17 2025 - Licensing: Proprietary - Benchmarks published at release: - SWE-Bench Pro: 49.6% - SWE-Bench Verified: 78% - Terminal-Bench 2.1: 58% - MCP Atlas: 62% - Toolathlon: 49.4% - Humanity's Last Exam (no tools): 33.7% - ARC-AGI-2: 33.6% - GPQA Diamond: 90.4% - OSWorld-Verified: 65.1% - Finance Agent v2: 42.6% - GDPval-AA: 1204 - CharXiv Reasoning: 80.3% - MMMU-Pro: 81.2% - Blueprint-Bench 2: 0% - MRCR v2 (8-needle) (128k average): 67.2% - MRCR v2 (8-needle) (1M pointwise): 22.1% - Arena Elo (Text): 1473 - Arena Elo (Code): 1437 **Q: When was Gemini 3.0 Flash released?** A: Gemini 3.0 Flash was released by Google on Wednesday, Dec 17 2025. **Q: Who made Gemini 3.0 Flash?** A: Gemini 3.0 Flash was built by Google. Builds the Gemini family of models through Google DeepMind. Integrates AI across Google products. **Q: What benchmark scores did Gemini 3.0 Flash get?** A: Gemini 3.0 Flash reports 18 tracked benchmark scores — SWE-Bench Pro: 49.6%; SWE-Bench Verified: 78%; Terminal-Bench 2.1: 58%; MCP Atlas: 62%; Toolathlon: 49.4%; Humanity's Last Exam (no tools): 33.7%; ARC-AGI-2: 33.6%; GPQA Diamond: 90.4%; OSWorld-Verified: 65.1%; Finance Agent v2: 42.6%; GDPval-AA: 1204; CharXiv Reasoning: 80.3%; MMMU-Pro: 81.2%; Blueprint-Bench 2: 0%; MRCR v2 (8-needle) (128k average): 67.2%; MRCR v2 (8-needle) (1M pointwise): 22.1%; Arena Elo (Text): 1473; Arena Elo (Code): 1437. Scores are the figures published at release by Google. **Q: Is Gemini 3.0 Flash open source?** A: No. Gemini 3.0 Flash is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Gemini 3.0 Flash?** A: Google's previous tracked release was Gemini 3.0 Pro on Nov 18 2025, 29 days earlier. It was followed by Gemini 3.1 Pro on Feb 19 2026. ### GPT-5.2 - URL: https://aireleasetracker.com/model/openai/gpt-5.2 - Developer: OpenAI - Release date: Thursday, Dec 11 2025 - Licensing: Proprietary - Benchmarks published at release: - BullshitBench v2: 38% - SWE-Bench Verified: 80% - ARC-AGI-2: 52.9% - GPQA Diamond: 92.4% - Arena Elo (Text): 1476 - Arena Elo (Code): 1419 GPT-5.2, released December 11, 2025, closed out OpenAI's 2025 with its strongest general model of the year: 92.4% on GPQA Diamond and 80.0% on SWE-Bench Verified, arriving less than a month after GPT-5.1 and directly answering Google's Gemini 3.0 Pro and Anthropic's Claude Opus 4.5 from the preceding weeks. It also scored 52.9% on ARC-AGI-2, at the time among the best published results on the abstract-reasoning benchmark. GPT-5.2 anchored OpenAI's lineup through the winter while the coding-focused GPT-5.3-Codex branch shipped in February 2026, and was ultimately succeeded by GPT-5.4 in March 2026. **Q: When was GPT-5.2 released?** A: GPT-5.2 was released by OpenAI on Thursday, Dec 11 2025. **Q: Who made GPT-5.2?** A: GPT-5.2 was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: What benchmark scores did GPT-5.2 get?** A: GPT-5.2 reports 6 tracked benchmark scores — BullshitBench v2: 38%; SWE-Bench Verified: 80%; ARC-AGI-2: 52.9%; GPQA Diamond: 92.4%; Arena Elo (Text): 1476; Arena Elo (Code): 1419. Scores are the figures published at release by OpenAI. **Q: Is GPT-5.2 open source?** A: No. GPT-5.2 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after GPT-5.2?** A: OpenAI's previous tracked release was GPT-5.1-Codex-Max on Nov 19 2025, 22 days earlier. It was followed by GPT-5.3-Codex on Feb 5 2026. ### Devstral 2 (123B) - URL: https://aireleasetracker.com/model/mistral/devstral-2-(123b) - Developer: Mistral - Release date: Tuesday, Dec 9 2025 - Licensing: Open Weight - Benchmarks: none published in this dataset **Q: When was Devstral 2 (123B) released?** A: Devstral 2 (123B) was released by Mistral on Tuesday, Dec 9 2025. **Q: Who made Devstral 2 (123B)?** A: Devstral 2 (123B) was built by Mistral. French AI company building open and commercial models. Founded in 2023 by former Meta and DeepMind researchers. **Q: Is Devstral 2 (123B) open source?** A: Partly. Devstral 2 (123B) is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Devstral 2 (123B)?** A: Mistral's previous tracked release was Mistral Large 3 on Dec 2 2025, 7 days earlier. It was followed by Devstral Small 2 (24B) on Dec 9 2025. ### Devstral Small 2 (24B) - URL: https://aireleasetracker.com/model/mistral/devstral-small-2-(24b) - Developer: Mistral - Release date: Tuesday, Dec 9 2025 - Licensing: Open Weight - Benchmarks: none published in this dataset **Q: When was Devstral Small 2 (24B) released?** A: Devstral Small 2 (24B) was released by Mistral on Tuesday, Dec 9 2025. **Q: Who made Devstral Small 2 (24B)?** A: Devstral Small 2 (24B) was built by Mistral. French AI company building open and commercial models. Founded in 2023 by former Meta and DeepMind researchers. **Q: Is Devstral Small 2 (24B) open source?** A: Partly. Devstral Small 2 (24B) is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Devstral Small 2 (24B)?** A: Mistral's previous tracked release was Devstral 2 (123B) on Dec 9 2025. It was followed by Mistral Small 4 on Mar 16 2026. ### Ministral 3 3B-25-12 - URL: https://aireleasetracker.com/model/mistral/ministral-3-3b-25-12 - Developer: Mistral - Release date: Tuesday, Dec 2 2025 - Licensing: Open Weight - Benchmarks: none published in this dataset **Q: When was Ministral 3 3B-25-12 released?** A: Ministral 3 3B-25-12 was released by Mistral on Tuesday, Dec 2 2025. **Q: Who made Ministral 3 3B-25-12?** A: Ministral 3 3B-25-12 was built by Mistral. French AI company building open and commercial models. Founded in 2023 by former Meta and DeepMind researchers. **Q: Is Ministral 3 3B-25-12 open source?** A: Partly. Ministral 3 3B-25-12 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Ministral 3 3B-25-12?** A: Mistral's previous tracked release was Mistral Small 3.2 on Jun 20 2025, 165 days earlier. It was followed by Ministral 3 8B-25-12 on Dec 2 2025. ### Ministral 3 8B-25-12 - URL: https://aireleasetracker.com/model/mistral/ministral-3-8b-25-12 - Developer: Mistral - Release date: Tuesday, Dec 2 2025 - Licensing: Open Weight - Benchmarks: none published in this dataset **Q: When was Ministral 3 8B-25-12 released?** A: Ministral 3 8B-25-12 was released by Mistral on Tuesday, Dec 2 2025. **Q: Who made Ministral 3 8B-25-12?** A: Ministral 3 8B-25-12 was built by Mistral. French AI company building open and commercial models. Founded in 2023 by former Meta and DeepMind researchers. **Q: Is Ministral 3 8B-25-12 open source?** A: Partly. Ministral 3 8B-25-12 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Ministral 3 8B-25-12?** A: Mistral's previous tracked release was Ministral 3 3B-25-12 on Dec 2 2025. It was followed by Ministral 3 14B-25-12 on Dec 2 2025. ### Ministral 3 14B-25-12 - URL: https://aireleasetracker.com/model/mistral/ministral-3-14b-25-12 - Developer: Mistral - Release date: Tuesday, Dec 2 2025 - Licensing: Open Weight - Benchmarks: none published in this dataset **Q: When was Ministral 3 14B-25-12 released?** A: Ministral 3 14B-25-12 was released by Mistral on Tuesday, Dec 2 2025. **Q: Who made Ministral 3 14B-25-12?** A: Ministral 3 14B-25-12 was built by Mistral. French AI company building open and commercial models. Founded in 2023 by former Meta and DeepMind researchers. **Q: Is Ministral 3 14B-25-12 open source?** A: Partly. Ministral 3 14B-25-12 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Ministral 3 14B-25-12?** A: Mistral's previous tracked release was Ministral 3 8B-25-12 on Dec 2 2025. It was followed by Mistral Large 3 on Dec 2 2025. ### Mistral Large 3 - URL: https://aireleasetracker.com/model/mistral/mistral-large-3 - Developer: Mistral - Release date: Tuesday, Dec 2 2025 - Licensing: Open Weight - Benchmarks published at release: - Arena Elo (Code): 1230 **Q: When was Mistral Large 3 released?** A: Mistral Large 3 was released by Mistral on Tuesday, Dec 2 2025. **Q: Who made Mistral Large 3?** A: Mistral Large 3 was built by Mistral. French AI company building open and commercial models. Founded in 2023 by former Meta and DeepMind researchers. **Q: What benchmark scores did Mistral Large 3 get?** A: Mistral Large 3 reports 1 tracked benchmark score — Arena Elo (Code): 1230. Scores are the figures published at release by Mistral. **Q: Is Mistral Large 3 open source?** A: Partly. Mistral Large 3 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Mistral Large 3?** A: Mistral's previous tracked release was Ministral 3 14B-25-12 on Dec 2 2025. It was followed by Devstral 2 (123B) on Dec 9 2025. ### DeepSeek-V3.2 - URL: https://aireleasetracker.com/model/deepseek/deepseek-v3.2 - Developer: DeepSeek - Release date: Monday, Dec 1 2025 - Licensing: Open Weight - Benchmarks published at release: - BullshitBench v2: 13% - Arena Elo (Code): 1361 **Q: When was DeepSeek-V3.2 released?** A: DeepSeek-V3.2 was released by DeepSeek on Monday, Dec 1 2025. **Q: Who made DeepSeek-V3.2?** A: DeepSeek-V3.2 was built by DeepSeek. Chinese AI lab known for efficient, open-weight models. Gained attention for strong performance at lower cost. **Q: What benchmark scores did DeepSeek-V3.2 get?** A: DeepSeek-V3.2 reports 2 tracked benchmark scores — BullshitBench v2: 13%; Arena Elo (Code): 1361. Scores are the figures published at release by DeepSeek. **Q: Is DeepSeek-V3.2 open source?** A: Partly. DeepSeek-V3.2 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after DeepSeek-V3.2?** A: DeepSeek's previous tracked release was DeepSeek V3.2 Exp on Sep 29 2025, 63 days earlier. It was followed by DeepSeek-V4-Pro on Apr 24 2026. ### Claude Opus 4.5 - URL: https://aireleasetracker.com/model/anthropic/claude-opus-4.5 - Developer: Anthropic - Release date: Monday, Nov 24 2025 - Licensing: Proprietary - Benchmarks published at release: - BullshitBench v2: 90% - SWE-Bench Verified: 80.9% - GPQA Diamond: 87% Claude Opus 4.5, released November 24, 2025, was the first model to break 80% on SWE-Bench Verified, scoring 80.9%, alongside 87.0% on GPQA Diamond. Anthropic paired the capability jump with a dramatic price cut for the Opus tier, making its flagship practical for everyday agentic coding rather than reserved for hard problems. The release capped an unusually fast 2025 cadence for Anthropic — Sonnet 4.5 in September, Haiku 4.5 in October, Opus 4.5 in November — and closed the year with Claude on top of most published coding evaluations. The Opus line continued with Opus 4.6 in February 2026 and Opus 4.7 in April 2026. **Q: When was Claude Opus 4.5 released?** A: Claude Opus 4.5 was released by Anthropic on Monday, Nov 24 2025. **Q: Who made Claude Opus 4.5?** A: Claude Opus 4.5 was built by Anthropic. AI safety company building the Claude family of models. Founded in 2021 by former OpenAI researchers. **Q: What benchmark scores did Claude Opus 4.5 get?** A: Claude Opus 4.5 reports 3 tracked benchmark scores — BullshitBench v2: 90%; SWE-Bench Verified: 80.9%; GPQA Diamond: 87%. Scores are the figures published at release by Anthropic. **Q: Is Claude Opus 4.5 open source?** A: No. Claude Opus 4.5 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Claude Opus 4.5?** A: Anthropic's previous tracked release was Claude Haiku 4.5 on Oct 15 2025, 40 days earlier. It was followed by Claude Opus 4.6 on Feb 5 2026. ### Grok 4.1 Fast - URL: https://aireleasetracker.com/model/xai/grok-4.1-fast - Developer: SpaceXAI - Release date: Wednesday, Nov 19 2025 - Licensing: Proprietary - Benchmarks published at release: - BullshitBench v2: 19% **Q: When was Grok 4.1 Fast released?** A: Grok 4.1 Fast was released by SpaceXAI on Wednesday, Nov 19 2025. **Q: Who made Grok 4.1 Fast?** A: Grok 4.1 Fast was built by SpaceXAI. Elon Musk's AI company building the Grok series of models. Founded in 2023 as xAI, now part of SpaceX. **Q: What benchmark scores did Grok 4.1 Fast get?** A: Grok 4.1 Fast reports 1 tracked benchmark score — BullshitBench v2: 19%. Scores are the figures published at release by SpaceXAI. **Q: Is Grok 4.1 Fast open source?** A: No. Grok 4.1 Fast is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Grok 4.1 Fast?** A: SpaceXAI's previous tracked release was Grok 4.1 on Nov 17 2025, 2 days earlier. It was followed by Grok 4.20 Beta on Feb 17 2026. ### GPT-5.1-Codex-Max - URL: https://aireleasetracker.com/model/openai/gpt-5.1-codex-max - Developer: OpenAI - Release date: Wednesday, Nov 19 2025 - Licensing: Proprietary - Benchmarks: none published in this dataset **Q: When was GPT-5.1-Codex-Max released?** A: GPT-5.1-Codex-Max was released by OpenAI on Wednesday, Nov 19 2025. **Q: Who made GPT-5.1-Codex-Max?** A: GPT-5.1-Codex-Max was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: Is GPT-5.1-Codex-Max open source?** A: No. GPT-5.1-Codex-Max is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after GPT-5.1-Codex-Max?** A: OpenAI's previous tracked release was GPT-5.1 on Nov 12 2025, 7 days earlier. It was followed by GPT-5.2 on Dec 11 2025. ### Gemini 3.0 Pro - URL: https://aireleasetracker.com/model/google/gemini-3.0-pro - Developer: Google - Release date: Tuesday, Nov 18 2025 - Licensing: Proprietary - Benchmarks published at release: - SWE-Bench Verified: 76.2% - Next.js Evals: 67% - ARC-AGI-2: 31.1% - GPQA Diamond: 91.9% - MMMU: 81% - Arena Elo (Text): 1486 - Arena Elo (Code): 1439 Gemini 3.0 Pro, released November 18, 2025, opened the most crowded six weeks in frontier-model history — Claude Opus 4.5 and GPT-5.2 both landed within a month of it. It scored 91.9% on GPQA Diamond and 81.0% on MMMU, the strongest multimodal-reasoning result of any model at release, with 76.2% on SWE-Bench Verified. The launch integrated Gemini deeply across Google Search's AI Mode and a new generation of agentic developer tooling, and the 3.0 line expanded with the fast Gemini 3.0 Flash in December 2025. Gemini 3.1 Pro succeeded it in February 2026 with a substantial jump in reasoning benchmarks. **Q: When was Gemini 3.0 Pro released?** A: Gemini 3.0 Pro was released by Google on Tuesday, Nov 18 2025. **Q: Who made Gemini 3.0 Pro?** A: Gemini 3.0 Pro was built by Google. Builds the Gemini family of models through Google DeepMind. Integrates AI across Google products. **Q: What benchmark scores did Gemini 3.0 Pro get?** A: Gemini 3.0 Pro reports 7 tracked benchmark scores — SWE-Bench Verified: 76.2%; Next.js Evals: 67%; ARC-AGI-2: 31.1%; GPQA Diamond: 91.9%; MMMU: 81%; Arena Elo (Text): 1486; Arena Elo (Code): 1439. Scores are the figures published at release by Google. **Q: Is Gemini 3.0 Pro open source?** A: No. Gemini 3.0 Pro is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Gemini 3.0 Pro?** A: Google's previous tracked release was Gemini 2.5 Flash-Lite on Jun 17 2025, 154 days earlier. It was followed by Gemini 3.0 Flash on Dec 17 2025. ### Grok 4.1 - URL: https://aireleasetracker.com/model/xai/grok-4.1 - Developer: SpaceXAI - Release date: Monday, Nov 17 2025 - Licensing: Proprietary - Benchmarks published at release: - Arena Elo (Text): 1466 - Arena Elo (Code): 1209 **Q: When was Grok 4.1 released?** A: Grok 4.1 was released by SpaceXAI on Monday, Nov 17 2025. **Q: Who made Grok 4.1?** A: Grok 4.1 was built by SpaceXAI. Elon Musk's AI company building the Grok series of models. Founded in 2023 as xAI, now part of SpaceX. **Q: What benchmark scores did Grok 4.1 get?** A: Grok 4.1 reports 2 tracked benchmark scores — Arena Elo (Text): 1466; Arena Elo (Code): 1209. Scores are the figures published at release by SpaceXAI. **Q: Is Grok 4.1 open source?** A: No. Grok 4.1 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Grok 4.1?** A: SpaceXAI's previous tracked release was Grok 4 Fast on Sep 19 2025, 59 days earlier. It was followed by Grok 4.1 Fast on Nov 19 2025. ### GPT-5.1 - URL: https://aireleasetracker.com/model/openai/gpt-5.1 - Developer: OpenAI - Release date: Wednesday, Nov 12 2025 - Licensing: Proprietary - Benchmarks published at release: - BullshitBench v2: 25% - SWE-Bench Verified: 76.3% - GPQA Diamond: 88.1% - MMMU: 76% - Arena Elo (Code): 1391 GPT-5.1, released November 12, 2025, was a usability-focused revision of GPT-5: a warmer, more conversational default tone, adaptive reasoning that spends thinking tokens only where they help, and Instant and Thinking variants replacing the opaque router of the original launch. The changes answered the most common criticisms of GPT-5's August debut. Capability moved too — 88.1% on GPQA Diamond and 76.3% on SWE-Bench Verified, with the coding-specialised GPT-5.1-Codex-Max following a week later. GPT-5.1 held the flagship spot for just under a month before GPT-5.2 arrived on December 11, 2025, in what was then the fastest flagship turnaround in OpenAI's history. **Q: When was GPT-5.1 released?** A: GPT-5.1 was released by OpenAI on Wednesday, Nov 12 2025. **Q: Who made GPT-5.1?** A: GPT-5.1 was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: What benchmark scores did GPT-5.1 get?** A: GPT-5.1 reports 5 tracked benchmark scores — BullshitBench v2: 25%; SWE-Bench Verified: 76.3%; GPQA Diamond: 88.1%; MMMU: 76%; Arena Elo (Code): 1391. Scores are the figures published at release by OpenAI. **Q: Is GPT-5.1 open source?** A: No. GPT-5.1 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after GPT-5.1?** A: OpenAI's previous tracked release was GPT-5-Codex-Mini on Nov 7 2025, 5 days earlier. It was followed by GPT-5.1-Codex-Max on Nov 19 2025. ### GPT-5-Codex-Mini - URL: https://aireleasetracker.com/model/openai/gpt-5-codex-mini - Developer: OpenAI - Release date: Friday, Nov 7 2025 - Licensing: Proprietary - Benchmarks: none published in this dataset **Q: When was GPT-5-Codex-Mini released?** A: GPT-5-Codex-Mini was released by OpenAI on Friday, Nov 7 2025. **Q: Who made GPT-5-Codex-Mini?** A: GPT-5-Codex-Mini was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: Is GPT-5-Codex-Mini open source?** A: No. GPT-5-Codex-Mini is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after GPT-5-Codex-Mini?** A: OpenAI's previous tracked release was GPT-5-Codex on Sep 15 2025, 53 days earlier. It was followed by GPT-5.1 on Nov 12 2025. ### Kimi K2 Thinking - URL: https://aireleasetracker.com/model/moonshot/kimi-k2-thinking - Developer: Moonshot AI - Release date: Thursday, Nov 6 2025 - Licensing: Open Weight - Parameters: 1T - Context window: 256k - Benchmarks published at release: - SWE-Bench Verified: 71.3% Kimi K2 Thinking, released November 6, 2025, added a full reasoning mode to Moonshot's trillion-parameter K2 line and was billed for long-horizon agentic work: the model was trained to sustain hundreds of consecutive tool calls without human intervention, interleaving thinking with action across a 256K context window. At release it posted the best agentic-search results of any open-weight model, including a leading score on Humanity's Last Exam with tools, and 71.3% on SWE-Bench Verified. Landing weeks before Claude Opus 4.5 and GPT-5.2, it made the strongest case yet that open weights could compete in the reasoning era. Kimi K2.5 followed in January 2026. **Q: When was Kimi K2 Thinking released?** A: Kimi K2 Thinking was released by Moonshot AI on Thursday, Nov 6 2025. **Q: Who made Kimi K2 Thinking?** A: Kimi K2 Thinking was built by Moonshot AI. Chinese AI company building the Kimi family of models. Founded in 2023, known for long-context and efficient MoE architectures. **Q: What benchmark scores did Kimi K2 Thinking get?** A: Kimi K2 Thinking reports 1 tracked benchmark score — SWE-Bench Verified: 71.3%. Scores are the figures published at release by Moonshot AI. **Q: What is the context window of Kimi K2 Thinking?** A: Kimi K2 Thinking has a context window of 256k. That is the maximum amount of input plus output the model can hold in a single request. **Q: How many parameters does Kimi K2 Thinking have?** A: Kimi K2 Thinking is reported at 1T parameters. **Q: Is Kimi K2 Thinking open source?** A: Partly. Kimi K2 Thinking is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Kimi K2 Thinking?** A: Moonshot AI's previous tracked release was Kimi Linear on Oct 30 2025, 7 days earlier. It was followed by Kimi K2.5 on Jan 27 2026. ### Kimi Linear - URL: https://aireleasetracker.com/model/moonshot/kimi-linear - Developer: Moonshot AI - Release date: Thursday, Oct 30 2025 - Licensing: Open Weight - Parameters: 48B - Benchmarks: none published in this dataset **Q: When was Kimi Linear released?** A: Kimi Linear was released by Moonshot AI on Thursday, Oct 30 2025. **Q: Who made Kimi Linear?** A: Kimi Linear was built by Moonshot AI. Chinese AI company building the Kimi family of models. Founded in 2023, known for long-context and efficient MoE architectures. **Q: How many parameters does Kimi Linear have?** A: Kimi Linear is reported at 48B parameters. **Q: Is Kimi Linear open source?** A: Partly. Kimi Linear is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Kimi Linear?** A: Moonshot AI's previous tracked release was Kimi K2 (0905) on Sep 5 2025, 55 days earlier. It was followed by Kimi K2 Thinking on Nov 6 2025. ### Composer - URL: https://aireleasetracker.com/model/cursor/composer - Developer: Cursor - Release date: Wednesday, Oct 29 2025 - Licensing: Proprietary - Benchmarks: none published in this dataset **Q: When was Composer released?** A: Composer was released by Cursor on Wednesday, Oct 29 2025. **Q: Who made Composer?** A: Composer was built by Cursor. **Q: Is Composer open source?** A: No. Composer is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Composer?** A: It was followed by Composer 1.5 on Feb 9 2026. ### Claude Haiku 4.5 - URL: https://aireleasetracker.com/model/anthropic/claude-haiku-4.5 - Developer: Anthropic - Release date: Wednesday, Oct 15 2025 - Licensing: Proprietary - Benchmarks published at release: - BullshitBench v2: 77% - SWE-Bench Verified: 73.3% - GPQA Diamond: 73% Claude Haiku 4.5, released October 15, 2025, brought near-frontier coding to Anthropic's smallest, fastest tier. It scored 73.3% on SWE-Bench Verified — essentially matching Claude Sonnet 4, the mid-tier flagship from just five months earlier — at roughly a third of the cost and more than twice the speed. That price-performance profile made Haiku 4.5 a popular choice for high-volume production workloads: sub-agents in orchestration systems, code review at scale, chat products with tight latency budgets. It also posted 73.0% on GPQA Diamond, territory that had belonged exclusively to flagship models a year earlier. **Q: When was Claude Haiku 4.5 released?** A: Claude Haiku 4.5 was released by Anthropic on Wednesday, Oct 15 2025. **Q: Who made Claude Haiku 4.5?** A: Claude Haiku 4.5 was built by Anthropic. AI safety company building the Claude family of models. Founded in 2021 by former OpenAI researchers. **Q: What benchmark scores did Claude Haiku 4.5 get?** A: Claude Haiku 4.5 reports 3 tracked benchmark scores — BullshitBench v2: 77%; SWE-Bench Verified: 73.3%; GPQA Diamond: 73%. Scores are the figures published at release by Anthropic. **Q: Is Claude Haiku 4.5 open source?** A: No. Claude Haiku 4.5 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Claude Haiku 4.5?** A: Anthropic's previous tracked release was Claude Sonnet 4.5 on Sep 29 2025, 16 days earlier. It was followed by Claude Opus 4.5 on Nov 24 2025. ### GLM-4.6 - URL: https://aireleasetracker.com/model/zai/glm-4.6 - Developer: Z.ai - Release date: Tuesday, Sep 30 2025 - Licensing: Open Weight - Parameters: 355B - Context window: 200k - Benchmarks published at release: - SWE-Bench Verified: 68% - Arena Elo (Code): 1339 GLM-4.6, released September 30, 2025, refined Z.ai's 355B open-weight flagship with a context window expanded from 128K to 200K tokens and markedly stronger real-world coding — 68.0% on SWE-Bench Verified, up eight points from GLM-4.5, with token efficiency improvements that cut the cost of agentic runs. Positioned aggressively against Claude Sonnet 4.5 (released the very next day) at a fraction of the price, GLM-4.6 became a staple budget option in coding agents like Claude Code-compatible harnesses and Cline. GLM-4.7 continued the line in December 2025 before the 744B GLM-5 generation arrived in February 2026. **Q: When was GLM-4.6 released?** A: GLM-4.6 was released by Z.ai on Tuesday, Sep 30 2025. **Q: Who made GLM-4.6?** A: GLM-4.6 was built by Z.ai. Chinese AI lab spun out of Tsinghua University (formerly Zhipu AI), building the open-weight GLM family. Rebranded internationally as Z.ai in 2025. **Q: What benchmark scores did GLM-4.6 get?** A: GLM-4.6 reports 2 tracked benchmark scores — SWE-Bench Verified: 68%; Arena Elo (Code): 1339. Scores are the figures published at release by Z.ai. **Q: What is the context window of GLM-4.6?** A: GLM-4.6 has a context window of 200k. That is the maximum amount of input plus output the model can hold in a single request. **Q: How many parameters does GLM-4.6 have?** A: GLM-4.6 is reported at 355B parameters. **Q: Is GLM-4.6 open source?** A: Partly. GLM-4.6 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after GLM-4.6?** A: Z.ai's previous tracked release was GLM-4.5-Air on Jul 28 2025, 64 days earlier. It was followed by GLM-4.7 on Dec 22 2025. ### Claude Sonnet 4.5 - URL: https://aireleasetracker.com/model/anthropic/claude-sonnet-4.5 - Developer: Anthropic - Release date: Monday, Sep 29 2025 - Licensing: Proprietary - Benchmarks published at release: - BullshitBench v2: 79% - SWE-Bench Verified: 77.2% - Next.js Evals: 50% - GPQA Diamond: 83.4% - MMMU: 68% Claude Sonnet 4.5, released September 29, 2025, was introduced by Anthropic as the best coding model in the world at the time. It reached 77.2% on SWE-Bench Verified and 83.4% on GPQA Diamond, and it was built for autonomy: Anthropic reported it could work productively on complex, multi-step tasks for many hours without losing the thread — the capability that underpinned the Claude Agent SDK, released the same day. Beyond raw scores, Sonnet 4.5 marked the point where the mid-tier Sonnet line overtook the previous flagship Opus 4.1 on software-engineering benchmarks while remaining far cheaper, a pattern that repeated across the industry as labs pushed agentic training into smaller models. It was succeeded at the top of the coding leaderboards by Claude Opus 4.5 in November 2025. **Q: When was Claude Sonnet 4.5 released?** A: Claude Sonnet 4.5 was released by Anthropic on Monday, Sep 29 2025. **Q: Who made Claude Sonnet 4.5?** A: Claude Sonnet 4.5 was built by Anthropic. AI safety company building the Claude family of models. Founded in 2021 by former OpenAI researchers. **Q: What benchmark scores did Claude Sonnet 4.5 get?** A: Claude Sonnet 4.5 reports 5 tracked benchmark scores — BullshitBench v2: 79%; SWE-Bench Verified: 77.2%; Next.js Evals: 50%; GPQA Diamond: 83.4%; MMMU: 68%. Scores are the figures published at release by Anthropic. **Q: Is Claude Sonnet 4.5 open source?** A: No. Claude Sonnet 4.5 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Claude Sonnet 4.5?** A: Anthropic's previous tracked release was Claude Opus 4.1 on Aug 5 2025, 55 days earlier. It was followed by Claude Haiku 4.5 on Oct 15 2025. ### DeepSeek V3.2 Exp - URL: https://aireleasetracker.com/model/deepseek/deepseek-v3.2-exp - Developer: DeepSeek - Release date: Monday, Sep 29 2025 - Licensing: Open Weight - Benchmarks published at release: - Arena Elo (Code): 1272 **Q: When was DeepSeek V3.2 Exp released?** A: DeepSeek V3.2 Exp was released by DeepSeek on Monday, Sep 29 2025. **Q: Who made DeepSeek V3.2 Exp?** A: DeepSeek V3.2 Exp was built by DeepSeek. Chinese AI lab known for efficient, open-weight models. Gained attention for strong performance at lower cost. **Q: What benchmark scores did DeepSeek V3.2 Exp get?** A: DeepSeek V3.2 Exp reports 1 tracked benchmark score — Arena Elo (Code): 1272. Scores are the figures published at release by DeepSeek. **Q: Is DeepSeek V3.2 Exp open source?** A: Partly. DeepSeek V3.2 Exp is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after DeepSeek V3.2 Exp?** A: DeepSeek's previous tracked release was DeepSeek V3.1 Terminus on Sep 22 2025, 7 days earlier. It was followed by DeepSeek-V3.2 on Dec 1 2025. ### Qwen3-Max - URL: https://aireleasetracker.com/model/qwen/qwen3-max - Developer: Qwen - Release date: Wednesday, Sep 24 2025 - Licensing: Proprietary - Parameters: 1T - Context window: 256k - Benchmarks: none published in this dataset Qwen3-Max, released September 24, 2025 at Alibaba's Apsara Conference, was the first Qwen model to cross a trillion parameters — and it stayed closed, available only through the Alibaba Cloud API with no weights published, continuing the pattern Qwen2.5-Max had set that January. For a lab whose reputation had been built almost entirely on open releases, keeping the largest models proprietary was a deliberate repositioning rather than an oversight. It was trained on roughly 36 trillion tokens and carried a 256K-token context window. The split it cemented — a proprietary Max tier at the top, open weights underneath — became the shape of Qwen's lineup from then on, with Qwen3.5-Plus, Qwen3.6-Plus and the Qwen3.7 pair all following the same commercial pattern while the numbered open releases continued in parallel. A reasoning variant, Qwen3-Max-Thinking, was added later. **Q: When was Qwen3-Max released?** A: Qwen3-Max was released by Qwen on Wednesday, Sep 24 2025. **Q: Who made Qwen3-Max?** A: Qwen3-Max was built by Qwen. Alibaba's AI lab, building the Qwen family. The most prolific publisher of open-weight models of any major lab, alongside a proprietary Max and Plus tier sold through Alibaba Cloud. **Q: What is the context window of Qwen3-Max?** A: Qwen3-Max has a context window of 256k. That is the maximum amount of input plus output the model can hold in a single request. **Q: How many parameters does Qwen3-Max have?** A: Qwen3-Max is reported at 1T parameters. **Q: Is Qwen3-Max open source?** A: No. Qwen3-Max is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Qwen3-Max?** A: Qwen's previous tracked release was Qwen3-Next on Sep 11 2025, 13 days earlier. It was followed by Qwen3-Coder-Next on Feb 3 2026. ### DeepSeek V3.1 Terminus - URL: https://aireleasetracker.com/model/deepseek/deepseek-v3.1-terminus - Developer: DeepSeek - Release date: Monday, Sep 22 2025 - Licensing: Open Weight - Benchmarks: none published in this dataset **Q: When was DeepSeek V3.1 Terminus released?** A: DeepSeek V3.1 Terminus was released by DeepSeek on Monday, Sep 22 2025. **Q: Who made DeepSeek V3.1 Terminus?** A: DeepSeek V3.1 Terminus was built by DeepSeek. Chinese AI lab known for efficient, open-weight models. Gained attention for strong performance at lower cost. **Q: Is DeepSeek V3.1 Terminus open source?** A: Partly. DeepSeek V3.1 Terminus is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after DeepSeek V3.1 Terminus?** A: DeepSeek's previous tracked release was DeepSeek V3.1 on Aug 21 2025, 32 days earlier. It was followed by DeepSeek V3.2 Exp on Sep 29 2025. ### Grok 4 Fast - URL: https://aireleasetracker.com/model/xai/grok-4-fast - Developer: SpaceXAI - Release date: Friday, Sep 19 2025 - Licensing: Proprietary - Benchmarks: none published in this dataset **Q: When was Grok 4 Fast released?** A: Grok 4 Fast was released by SpaceXAI on Friday, Sep 19 2025. **Q: Who made Grok 4 Fast?** A: Grok 4 Fast was built by SpaceXAI. Elon Musk's AI company building the Grok series of models. Founded in 2023 as xAI, now part of SpaceX. **Q: Is Grok 4 Fast open source?** A: No. Grok 4 Fast is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Grok 4 Fast?** A: SpaceXAI's previous tracked release was Grok Code Fast 1 on Aug 28 2025, 22 days earlier. It was followed by Grok 4.1 on Nov 17 2025. ### GPT-5-Codex - URL: https://aireleasetracker.com/model/openai/gpt-5-codex - Developer: OpenAI - Release date: Monday, Sep 15 2025 - Licensing: Proprietary - Benchmarks published at release: - BullshitBench v2: 39% **Q: When was GPT-5-Codex released?** A: GPT-5-Codex was released by OpenAI on Monday, Sep 15 2025. **Q: Who made GPT-5-Codex?** A: GPT-5-Codex was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: What benchmark scores did GPT-5-Codex get?** A: GPT-5-Codex reports 1 tracked benchmark score — BullshitBench v2: 39%. Scores are the figures published at release by OpenAI. **Q: Is GPT-5-Codex open source?** A: No. GPT-5-Codex is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after GPT-5-Codex?** A: OpenAI's previous tracked release was GPT-5 Pro on Aug 7 2025, 39 days earlier. It was followed by GPT-5-Codex-Mini on Nov 7 2025. ### Qwen3-Next - URL: https://aireleasetracker.com/model/qwen/qwen3-next - Developer: Qwen - Release date: Thursday, Sep 11 2025 - Licensing: Open Weight - Parameters: 80B - Context window: 256k - Benchmarks: none published in this dataset **Q: When was Qwen3-Next released?** A: Qwen3-Next was released by Qwen on Thursday, Sep 11 2025. **Q: Who made Qwen3-Next?** A: Qwen3-Next was built by Qwen. Alibaba's AI lab, building the Qwen family. The most prolific publisher of open-weight models of any major lab, alongside a proprietary Max and Plus tier sold through Alibaba Cloud. **Q: What is the context window of Qwen3-Next?** A: Qwen3-Next has a context window of 256k. That is the maximum amount of input plus output the model can hold in a single request. **Q: How many parameters does Qwen3-Next have?** A: Qwen3-Next is reported at 80B parameters. **Q: Is Qwen3-Next open source?** A: Partly. Qwen3-Next is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Qwen3-Next?** A: Qwen's previous tracked release was Qwen3-Coder on Jul 22 2025, 51 days earlier. It was followed by Qwen3-Max on Sep 24 2025. ### Kimi K2 (0905) - URL: https://aireleasetracker.com/model/moonshot/kimi-k2-(0905) - Developer: Moonshot AI - Release date: Friday, Sep 5 2025 - Licensing: Open Weight - Parameters: 1T - Context window: 256k - Benchmarks published at release: - SWE-Bench Verified: 65.8% - GPQA Diamond: 75.1% **Q: When was Kimi K2 (0905) released?** A: Kimi K2 (0905) was released by Moonshot AI on Friday, Sep 5 2025. **Q: Who made Kimi K2 (0905)?** A: Kimi K2 (0905) was built by Moonshot AI. Chinese AI company building the Kimi family of models. Founded in 2023, known for long-context and efficient MoE architectures. **Q: What benchmark scores did Kimi K2 (0905) get?** A: Kimi K2 (0905) reports 2 tracked benchmark scores — SWE-Bench Verified: 65.8%; GPQA Diamond: 75.1%. Scores are the figures published at release by Moonshot AI. **Q: What is the context window of Kimi K2 (0905)?** A: Kimi K2 (0905) has a context window of 256k. That is the maximum amount of input plus output the model can hold in a single request. **Q: How many parameters does Kimi K2 (0905) have?** A: Kimi K2 (0905) is reported at 1T parameters. **Q: Is Kimi K2 (0905) open source?** A: Partly. Kimi K2 (0905) is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Kimi K2 (0905)?** A: Moonshot AI's previous tracked release was Kimi K2 on Jul 11 2025, 56 days earlier. It was followed by Kimi Linear on Oct 30 2025. ### Grok Code Fast 1 - URL: https://aireleasetracker.com/model/xai/grok-code-fast-1 - Developer: SpaceXAI - Release date: Thursday, Aug 28 2025 - Licensing: Proprietary - Benchmarks published at release: - Arena Elo (Code): 1163 **Q: When was Grok Code Fast 1 released?** A: Grok Code Fast 1 was released by SpaceXAI on Thursday, Aug 28 2025. **Q: Who made Grok Code Fast 1?** A: Grok Code Fast 1 was built by SpaceXAI. Elon Musk's AI company building the Grok series of models. Founded in 2023 as xAI, now part of SpaceX. **Q: What benchmark scores did Grok Code Fast 1 get?** A: Grok Code Fast 1 reports 1 tracked benchmark score — Arena Elo (Code): 1163. Scores are the figures published at release by SpaceXAI. **Q: Is Grok Code Fast 1 open source?** A: No. Grok Code Fast 1 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Grok Code Fast 1?** A: SpaceXAI's previous tracked release was Grok 2.5 on Aug 23 2025, 5 days earlier. It was followed by Grok 4 Fast on Sep 19 2025. ### Grok 2.5 - URL: https://aireleasetracker.com/model/xai/grok-2.5 - Developer: SpaceXAI - Release date: Saturday, Aug 23 2025 - Licensing: Proprietary - Benchmarks: none published in this dataset **Q: When was Grok 2.5 released?** A: Grok 2.5 was released by SpaceXAI on Saturday, Aug 23 2025. **Q: Who made Grok 2.5?** A: Grok 2.5 was built by SpaceXAI. Elon Musk's AI company building the Grok series of models. Founded in 2023 as xAI, now part of SpaceX. **Q: Is Grok 2.5 open source?** A: No. Grok 2.5 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Grok 2.5?** A: SpaceXAI's previous tracked release was Grok 4 Heavy on Jul 9 2025, 45 days earlier. It was followed by Grok Code Fast 1 on Aug 28 2025. ### DeepSeek V3.1 - URL: https://aireleasetracker.com/model/deepseek/deepseek-v3.1 - Developer: DeepSeek - Release date: Thursday, Aug 21 2025 - Licensing: Open Weight - Benchmarks: none published in this dataset DeepSeek V3.1, released August 21, 2025, merged the V3 and R1 lines into a single hybrid model with switchable thinking and non-thinking modes, tuned for the agent era: stronger tool calling, better multi-step task execution, and faster reasoning than R1 at similar quality. As with every DeepSeek release, the weights were open. It marked DeepSeek's shift from chasing benchmark parity toward practical agentic use, and was refined by V3.1 Terminus in September 2025 before the experimental V3.2 introduced sparse attention to cut long-context costs. The line ultimately led to the V4 generation in April 2026. **Q: When was DeepSeek V3.1 released?** A: DeepSeek V3.1 was released by DeepSeek on Thursday, Aug 21 2025. **Q: Who made DeepSeek V3.1?** A: DeepSeek V3.1 was built by DeepSeek. Chinese AI lab known for efficient, open-weight models. Gained attention for strong performance at lower cost. **Q: Is DeepSeek V3.1 open source?** A: Partly. DeepSeek V3.1 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after DeepSeek V3.1?** A: DeepSeek's previous tracked release was DeepSeek R1-0528 on May 28 2025, 85 days earlier. It was followed by DeepSeek V3.1 Terminus on Sep 22 2025. ### GPT-5 - URL: https://aireleasetracker.com/model/openai/gpt-5 - Developer: OpenAI - Release date: Thursday, Aug 7 2025 - Licensing: Proprietary - Benchmarks published at release: - BullshitBench v2: 21% - Arena Elo (Code): 1419 GPT-5, released August 7, 2025, collapsed OpenAI's confusing two-track lineup — GPT-4.1 for speed, o3 for reasoning — into a single system with a router that decides per request how much thinking a question deserves. It launched simultaneously in mini, nano, and Pro variants and became the default for ChatGPT's hundreds of millions of users, including the free tier. The launch drew a famously mixed reception: benchmark gains over o3 were real but incremental, and the router initially frustrated users who wanted explicit model control. OpenAI iterated quickly — GPT-5.1 in November 2025 reworked the model's personality and adaptive reasoning, and the GPT-5.x line has carried OpenAI's flagship branding ever since. **Q: When was GPT-5 released?** A: GPT-5 was released by OpenAI on Thursday, Aug 7 2025. **Q: Who made GPT-5?** A: GPT-5 was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: What benchmark scores did GPT-5 get?** A: GPT-5 reports 2 tracked benchmark scores — BullshitBench v2: 21%; Arena Elo (Code): 1419. Scores are the figures published at release by OpenAI. **Q: Is GPT-5 open source?** A: No. GPT-5 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after GPT-5?** A: OpenAI's previous tracked release was gpt-oss-20b on Aug 5 2025, 2 days earlier. It was followed by GPT-5 mini on Aug 7 2025. ### GPT-5 mini - URL: https://aireleasetracker.com/model/openai/gpt-5-mini - Developer: OpenAI - Release date: Thursday, Aug 7 2025 - Licensing: Proprietary - Benchmarks: none published in this dataset **Q: When was GPT-5 mini released?** A: GPT-5 mini was released by OpenAI on Thursday, Aug 7 2025. **Q: Who made GPT-5 mini?** A: GPT-5 mini was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: Is GPT-5 mini open source?** A: No. GPT-5 mini is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after GPT-5 mini?** A: OpenAI's previous tracked release was GPT-5 on Aug 7 2025. It was followed by GPT-5 nano on Aug 7 2025. ### GPT-5 nano - URL: https://aireleasetracker.com/model/openai/gpt-5-nano - Developer: OpenAI - Release date: Thursday, Aug 7 2025 - Licensing: Proprietary - Benchmarks: none published in this dataset **Q: When was GPT-5 nano released?** A: GPT-5 nano was released by OpenAI on Thursday, Aug 7 2025. **Q: Who made GPT-5 nano?** A: GPT-5 nano was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: Is GPT-5 nano open source?** A: No. GPT-5 nano is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after GPT-5 nano?** A: OpenAI's previous tracked release was GPT-5 mini on Aug 7 2025. It was followed by GPT-5 Pro on Aug 7 2025. ### GPT-5 Pro - URL: https://aireleasetracker.com/model/openai/gpt-5-pro - Developer: OpenAI - Release date: Thursday, Aug 7 2025 - Licensing: Proprietary - Benchmarks: none published in this dataset **Q: When was GPT-5 Pro released?** A: GPT-5 Pro was released by OpenAI on Thursday, Aug 7 2025. **Q: Who made GPT-5 Pro?** A: GPT-5 Pro was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: Is GPT-5 Pro open source?** A: No. GPT-5 Pro is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after GPT-5 Pro?** A: OpenAI's previous tracked release was GPT-5 nano on Aug 7 2025. It was followed by GPT-5-Codex on Sep 15 2025. ### Claude Opus 4.1 - URL: https://aireleasetracker.com/model/anthropic/claude-opus-4.1 - Developer: Anthropic - Release date: Tuesday, Aug 5 2025 - Licensing: Proprietary - Benchmarks published at release: - BullshitBench v2: 43% - SWE-Bench Verified: 74.5% - GPQA Diamond: 80.9% **Q: When was Claude Opus 4.1 released?** A: Claude Opus 4.1 was released by Anthropic on Tuesday, Aug 5 2025. **Q: Who made Claude Opus 4.1?** A: Claude Opus 4.1 was built by Anthropic. AI safety company building the Claude family of models. Founded in 2021 by former OpenAI researchers. **Q: What benchmark scores did Claude Opus 4.1 get?** A: Claude Opus 4.1 reports 3 tracked benchmark scores — BullshitBench v2: 43%; SWE-Bench Verified: 74.5%; GPQA Diamond: 80.9%. Scores are the figures published at release by Anthropic. **Q: Is Claude Opus 4.1 open source?** A: No. Claude Opus 4.1 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Claude Opus 4.1?** A: Anthropic's previous tracked release was Claude Opus 4 on May 22 2025, 75 days earlier. It was followed by Claude Sonnet 4.5 on Sep 29 2025. ### gpt-oss-120b - URL: https://aireleasetracker.com/model/openai/gpt-oss-120b - Developer: OpenAI - Release date: Tuesday, Aug 5 2025 - Licensing: Open Weight - Parameters: 117B - Context window: 128k - Benchmarks published at release: - BullshitBench v2: 11% - SWE-Bench Verified: 62.4% - Humanity's Last Exam (no tools): 14.9% - Humanity's Last Exam (with tools): 19% - GPQA Diamond: 80.1% - MMLU: 90% gpt-oss-120b, released August 5, 2025 alongside the smaller gpt-oss-20b, was OpenAI's first open-weight language model since GPT-2 in 2019 — six years in which the company that popularised the term "open" shipped nothing downloadable. Both models came under Apache 2.0 with no usage restrictions and no commercial gate. The 120B is a mixture-of-experts design with about 5.1B parameters active per token, sized to fit on a single 80GB accelerator, while the 21B sibling was built to run on a high-end laptop. The launch figures put it close to OpenAI's own mid-tier closed models: 80.1% on GPQA Diamond and 62.4% on SWE-Bench Verified, 90.0% on MMLU, and 14.9% on Humanity's Last Exam without tools, rising to 19.0% with them. OpenAI framed the release as a response to the Chinese open-weight wave — DeepSeek R1 and Kimi K2 had spent the preceding months setting the pace for freely downloadable models — and shipped it two days before GPT-5, which took the attention. **Q: When was gpt-oss-120b released?** A: gpt-oss-120b was released by OpenAI on Tuesday, Aug 5 2025. **Q: Who made gpt-oss-120b?** A: gpt-oss-120b was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: What benchmark scores did gpt-oss-120b get?** A: gpt-oss-120b reports 6 tracked benchmark scores — BullshitBench v2: 11%; SWE-Bench Verified: 62.4%; Humanity's Last Exam (no tools): 14.9%; Humanity's Last Exam (with tools): 19%; GPQA Diamond: 80.1%; MMLU: 90%. Scores are the figures published at release by OpenAI. **Q: What is the context window of gpt-oss-120b?** A: gpt-oss-120b has a context window of 128k. That is the maximum amount of input plus output the model can hold in a single request. **Q: How many parameters does gpt-oss-120b have?** A: gpt-oss-120b is reported at 117B parameters. **Q: Is gpt-oss-120b open source?** A: Partly. gpt-oss-120b is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after gpt-oss-120b?** A: OpenAI's previous tracked release was o3-pro on Jun 10 2025, 56 days earlier. It was followed by gpt-oss-20b on Aug 5 2025. ### gpt-oss-20b - URL: https://aireleasetracker.com/model/openai/gpt-oss-20b - Developer: OpenAI - Release date: Tuesday, Aug 5 2025 - Licensing: Open Weight - Parameters: 21B - Context window: 128k - Benchmarks published at release: - SWE-Bench Verified: 60.7% - Humanity's Last Exam (no tools): 10.9% - Humanity's Last Exam (with tools): 17.3% - GPQA Diamond: 71.5% - MMLU: 85.3% **Q: When was gpt-oss-20b released?** A: gpt-oss-20b was released by OpenAI on Tuesday, Aug 5 2025. **Q: Who made gpt-oss-20b?** A: gpt-oss-20b was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: What benchmark scores did gpt-oss-20b get?** A: gpt-oss-20b reports 5 tracked benchmark scores — SWE-Bench Verified: 60.7%; Humanity's Last Exam (no tools): 10.9%; Humanity's Last Exam (with tools): 17.3%; GPQA Diamond: 71.5%; MMLU: 85.3%. Scores are the figures published at release by OpenAI. **Q: What is the context window of gpt-oss-20b?** A: gpt-oss-20b has a context window of 128k. That is the maximum amount of input plus output the model can hold in a single request. **Q: How many parameters does gpt-oss-20b have?** A: gpt-oss-20b is reported at 21B parameters. **Q: Is gpt-oss-20b open source?** A: Partly. gpt-oss-20b is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after gpt-oss-20b?** A: OpenAI's previous tracked release was gpt-oss-120b on Aug 5 2025. It was followed by GPT-5 on Aug 7 2025. ### GLM-4.5 - URL: https://aireleasetracker.com/model/zai/glm-4.5 - Developer: Z.ai - Release date: Monday, Jul 28 2025 - Licensing: Open Weight - Parameters: 355B - Context window: 128k - Benchmarks published at release: - BullshitBench v2: 8% - SWE-Bench Verified: 64.2% - GPQA Diamond: 79.1% GLM-4.5, released July 28, 2025 alongside the lighter GLM-4.5-Air, was Z.ai's entry into the open-weight frontier race: a 355B-parameter mixture-of-experts model designed explicitly for agentic tasks, unifying reasoning, coding, and tool use with switchable thinking modes. It scored 79.1% on GPQA and 64.2% on SWE-Bench Verified. Released the same month as Kimi K2, it was part of the mid-2025 wave of Chinese open-weight agentic models that reset expectations for what freely downloadable weights could do. Its low-cost API and coding-agent integrations won it a significant developer following outside China, and GLM-4.6 followed nine weeks later. **Q: When was GLM-4.5 released?** A: GLM-4.5 was released by Z.ai on Monday, Jul 28 2025. **Q: Who made GLM-4.5?** A: GLM-4.5 was built by Z.ai. Chinese AI lab spun out of Tsinghua University (formerly Zhipu AI), building the open-weight GLM family. Rebranded internationally as Z.ai in 2025. **Q: What benchmark scores did GLM-4.5 get?** A: GLM-4.5 reports 3 tracked benchmark scores — BullshitBench v2: 8%; SWE-Bench Verified: 64.2%; GPQA Diamond: 79.1%. Scores are the figures published at release by Z.ai. **Q: What is the context window of GLM-4.5?** A: GLM-4.5 has a context window of 128k. That is the maximum amount of input plus output the model can hold in a single request. **Q: How many parameters does GLM-4.5 have?** A: GLM-4.5 is reported at 355B parameters. **Q: Is GLM-4.5 open source?** A: Partly. GLM-4.5 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after GLM-4.5?** A: Z.ai's previous tracked release was GLM-4-Plus on Aug 29 2024, 333 days earlier. It was followed by GLM-4.5-Air on Jul 28 2025. ### GLM-4.5-Air - URL: https://aireleasetracker.com/model/zai/glm-4.5-air - Developer: Z.ai - Release date: Monday, Jul 28 2025 - Licensing: Open Weight - Parameters: 106B - Benchmarks published at release: - SWE-Bench Verified: 59.8% **Q: When was GLM-4.5-Air released?** A: GLM-4.5-Air was released by Z.ai on Monday, Jul 28 2025. **Q: Who made GLM-4.5-Air?** A: GLM-4.5-Air was built by Z.ai. Chinese AI lab spun out of Tsinghua University (formerly Zhipu AI), building the open-weight GLM family. Rebranded internationally as Z.ai in 2025. **Q: What benchmark scores did GLM-4.5-Air get?** A: GLM-4.5-Air reports 1 tracked benchmark score — SWE-Bench Verified: 59.8%. Scores are the figures published at release by Z.ai. **Q: How many parameters does GLM-4.5-Air have?** A: GLM-4.5-Air is reported at 106B parameters. **Q: Is GLM-4.5-Air open source?** A: Partly. GLM-4.5-Air is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after GLM-4.5-Air?** A: Z.ai's previous tracked release was GLM-4.5 on Jul 28 2025. It was followed by GLM-4.6 on Sep 30 2025. ### Qwen3-Coder - URL: https://aireleasetracker.com/model/qwen/qwen3-coder - Developer: Qwen - Release date: Tuesday, Jul 22 2025 - Licensing: Open Weight - Parameters: 480B - Context window: 256k - Benchmarks published at release: - BullshitBench v2: 20% Qwen3-Coder, released July 22, 2025, was Alibaba's largest open-weight model to that point: a 480B-parameter mixture-of-experts model with 35B active, built specifically for agentic coding rather than code completion. It shipped with a 256K-token native context window, extensible to roughly 1M through positional extrapolation, so a whole repository could sit inside a single prompt. It arrived alongside Qwen Code, a command-line agent forked from Google's Gemini CLI, which signalled what the model was for — long multi-step tool-using sessions rather than single-shot completions. Released the same month as Kimi K2 and GLM-4.5, it formed part of a mid-2025 cluster of Chinese open-weight agentic models that reset what freely downloadable weights were expected to do, and a smaller Qwen3-Coder-Flash followed for latency-sensitive work. **Q: When was Qwen3-Coder released?** A: Qwen3-Coder was released by Qwen on Tuesday, Jul 22 2025. **Q: Who made Qwen3-Coder?** A: Qwen3-Coder was built by Qwen. Alibaba's AI lab, building the Qwen family. The most prolific publisher of open-weight models of any major lab, alongside a proprietary Max and Plus tier sold through Alibaba Cloud. **Q: What benchmark scores did Qwen3-Coder get?** A: Qwen3-Coder reports 1 tracked benchmark score — BullshitBench v2: 20%. Scores are the figures published at release by Qwen. **Q: What is the context window of Qwen3-Coder?** A: Qwen3-Coder has a context window of 256k. That is the maximum amount of input plus output the model can hold in a single request. **Q: How many parameters does Qwen3-Coder have?** A: Qwen3-Coder is reported at 480B parameters. **Q: Is Qwen3-Coder open source?** A: Partly. Qwen3-Coder is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Qwen3-Coder?** A: Qwen's previous tracked release was Qwen3 on Apr 29 2025, 84 days earlier. It was followed by Qwen3-Next on Sep 11 2025. ### Kimi K2 - URL: https://aireleasetracker.com/model/moonshot/kimi-k2 - Developer: Moonshot AI - Release date: Friday, Jul 11 2025 - Licensing: Open Weight - Parameters: 1T - Context window: 128k - Benchmarks published at release: - BullshitBench v2: 10% - SWE-Bench Verified: 65.8% - GPQA Diamond: 75.1% Kimi K2, released July 11, 2025, was Moonshot AI's open-weight breakout: a trillion-parameter mixture-of-experts model with about 32B active per token, trained with the novel MuonClip optimizer and tuned specifically for agentic tool use rather than chat. At release it was the largest open-weight model ever published. It scored 65.8% on SWE-Bench Verified and 75.1% on GPQA Diamond — the strongest open-weight coding results of mid-2025 — and its low API pricing made it a popular Claude alternative in agent frameworks. The K2 (0905) refresh doubled the context window to 256K in September 2025, and K2 Thinking added reasoning that November. **Q: When was Kimi K2 released?** A: Kimi K2 was released by Moonshot AI on Friday, Jul 11 2025. **Q: Who made Kimi K2?** A: Kimi K2 was built by Moonshot AI. Chinese AI company building the Kimi family of models. Founded in 2023, known for long-context and efficient MoE architectures. **Q: What benchmark scores did Kimi K2 get?** A: Kimi K2 reports 3 tracked benchmark scores — BullshitBench v2: 10%; SWE-Bench Verified: 65.8%; GPQA Diamond: 75.1%. Scores are the figures published at release by Moonshot AI. **Q: What is the context window of Kimi K2?** A: Kimi K2 has a context window of 128k. That is the maximum amount of input plus output the model can hold in a single request. **Q: How many parameters does Kimi K2 have?** A: Kimi K2 is reported at 1T parameters. **Q: Is Kimi K2 open source?** A: Partly. Kimi K2 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Kimi K2?** A: Moonshot AI's previous tracked release was Kimi-Dev on Jun 17 2025, 24 days earlier. It was followed by Kimi K2 (0905) on Sep 5 2025. ### Grok 4 - URL: https://aireleasetracker.com/model/xai/grok-4 - Developer: SpaceXAI - Release date: Wednesday, Jul 9 2025 - Licensing: Proprietary - Benchmarks: none published in this dataset Grok 4, released July 9, 2025, was xAI's reasoning-first flagship, trained with large-scale reinforcement learning on the Colossus cluster. It launched with claims of leading scores on Humanity's Last Exam and ARC-AGI-2, and shipped alongside Grok 4 Heavy, a multi-agent variant that runs several instances in parallel and compares their answers. The launch cemented xAI's pattern of aggressive benchmark competition and rapid iteration: Grok 4 Fast followed in September 2025, the coding-focused Grok Code Fast 1 in August, and Grok 4.1 in November 2025. The line continued with Grok 4.5 in July 2026 and Grok 4.6 in August 2026. **Q: When was Grok 4 released?** A: Grok 4 was released by SpaceXAI on Wednesday, Jul 9 2025. **Q: Who made Grok 4?** A: Grok 4 was built by SpaceXAI. Elon Musk's AI company building the Grok series of models. Founded in 2023 as xAI, now part of SpaceX. **Q: Is Grok 4 open source?** A: No. Grok 4 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Grok 4?** A: SpaceXAI's previous tracked release was Grok 3 mini on Feb 17 2025, 142 days earlier. It was followed by Grok 4 Heavy on Jul 9 2025. ### Grok 4 Heavy - URL: https://aireleasetracker.com/model/xai/grok-4-heavy - Developer: SpaceXAI - Release date: Wednesday, Jul 9 2025 - Licensing: Proprietary - Benchmarks: none published in this dataset **Q: When was Grok 4 Heavy released?** A: Grok 4 Heavy was released by SpaceXAI on Wednesday, Jul 9 2025. **Q: Who made Grok 4 Heavy?** A: Grok 4 Heavy was built by SpaceXAI. Elon Musk's AI company building the Grok series of models. Founded in 2023 as xAI, now part of SpaceX. **Q: Is Grok 4 Heavy open source?** A: No. Grok 4 Heavy is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Grok 4 Heavy?** A: SpaceXAI's previous tracked release was Grok 4 on Jul 9 2025. It was followed by Grok 2.5 on Aug 23 2025. ### Mistral Small 3.2 - URL: https://aireleasetracker.com/model/mistral/mistral-small-3.2 - Developer: Mistral - Release date: Friday, Jun 20 2025 - Licensing: Open Weight - Parameters: 24B - Benchmarks: none published in this dataset **Q: When was Mistral Small 3.2 released?** A: Mistral Small 3.2 was released by Mistral on Friday, Jun 20 2025. **Q: Who made Mistral Small 3.2?** A: Mistral Small 3.2 was built by Mistral. French AI company building open and commercial models. Founded in 2023 by former Meta and DeepMind researchers. **Q: How many parameters does Mistral Small 3.2 have?** A: Mistral Small 3.2 is reported at 24B parameters. **Q: Is Mistral Small 3.2 open source?** A: Partly. Mistral Small 3.2 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Mistral Small 3.2?** A: Mistral's previous tracked release was Magistral Medium on Jun 10 2025, 10 days earlier. It was followed by Ministral 3 3B-25-12 on Dec 2 2025. ### Gemini 2.5 Flash-Lite - URL: https://aireleasetracker.com/model/google/gemini-2.5-flash-lite - Developer: Google - Release date: Tuesday, Jun 17 2025 - Licensing: Proprietary - Benchmarks: none published in this dataset **Q: When was Gemini 2.5 Flash-Lite released?** A: Gemini 2.5 Flash-Lite was released by Google on Tuesday, Jun 17 2025. **Q: Who made Gemini 2.5 Flash-Lite?** A: Gemini 2.5 Flash-Lite was built by Google. Builds the Gemini family of models through Google DeepMind. Integrates AI across Google products. **Q: Is Gemini 2.5 Flash-Lite open source?** A: No. Gemini 2.5 Flash-Lite is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Gemini 2.5 Flash-Lite?** A: Google's previous tracked release was Gemini 2.5 Flash on Apr 17 2025, 61 days earlier. It was followed by Gemini 3.0 Pro on Nov 18 2025. ### Kimi-Dev - URL: https://aireleasetracker.com/model/moonshot/kimi-dev - Developer: Moonshot AI - Release date: Tuesday, Jun 17 2025 - Licensing: Open Weight - Parameters: 72B - Benchmarks: none published in this dataset **Q: When was Kimi-Dev released?** A: Kimi-Dev was released by Moonshot AI on Tuesday, Jun 17 2025. **Q: Who made Kimi-Dev?** A: Kimi-Dev was built by Moonshot AI. Chinese AI company building the Kimi family of models. Founded in 2023, known for long-context and efficient MoE architectures. **Q: How many parameters does Kimi-Dev have?** A: Kimi-Dev is reported at 72B parameters. **Q: Is Kimi-Dev open source?** A: Partly. Kimi-Dev is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Kimi-Dev?** A: Moonshot AI's previous tracked release was Kimi-VL on Apr 10 2025, 68 days earlier. It was followed by Kimi K2 on Jul 11 2025. ### o3-pro - URL: https://aireleasetracker.com/model/openai/o3-pro - Developer: OpenAI - Release date: Tuesday, Jun 10 2025 - Licensing: Proprietary - Benchmarks: none published in this dataset **Q: When was o3-pro released?** A: o3-pro was released by OpenAI on Tuesday, Jun 10 2025. **Q: Who made o3-pro?** A: o3-pro was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: Is o3-pro open source?** A: No. o3-pro is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after o3-pro?** A: OpenAI's previous tracked release was o4-mini-high on Apr 16 2025, 55 days earlier. It was followed by gpt-oss-120b on Aug 5 2025. ### Magistral Small - URL: https://aireleasetracker.com/model/mistral/magistral-small - Developer: Mistral - Release date: Tuesday, Jun 10 2025 - Licensing: Proprietary - Benchmarks: none published in this dataset **Q: When was Magistral Small released?** A: Magistral Small was released by Mistral on Tuesday, Jun 10 2025. **Q: Who made Magistral Small?** A: Magistral Small was built by Mistral. French AI company building open and commercial models. Founded in 2023 by former Meta and DeepMind researchers. **Q: Is Magistral Small open source?** A: No. Magistral Small is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Magistral Small?** A: Mistral's previous tracked release was Devstral Small on May 21 2025, 20 days earlier. It was followed by Magistral Medium on Jun 10 2025. ### Magistral Medium - URL: https://aireleasetracker.com/model/mistral/magistral-medium - Developer: Mistral - Release date: Tuesday, Jun 10 2025 - Licensing: Proprietary - Benchmarks: none published in this dataset **Q: When was Magistral Medium released?** A: Magistral Medium was released by Mistral on Tuesday, Jun 10 2025. **Q: Who made Magistral Medium?** A: Magistral Medium was built by Mistral. French AI company building open and commercial models. Founded in 2023 by former Meta and DeepMind researchers. **Q: Is Magistral Medium open source?** A: No. Magistral Medium is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Magistral Medium?** A: Mistral's previous tracked release was Magistral Small on Jun 10 2025. It was followed by Mistral Small 3.2 on Jun 20 2025. ### DeepSeek R1-0528 - URL: https://aireleasetracker.com/model/deepseek/deepseek-r1-0528 - Developer: DeepSeek - Release date: Wednesday, May 28 2025 - Licensing: Open Weight - Benchmarks published at release: - BullshitBench v2: 8% **Q: When was DeepSeek R1-0528 released?** A: DeepSeek R1-0528 was released by DeepSeek on Wednesday, May 28 2025. **Q: Who made DeepSeek R1-0528?** A: DeepSeek R1-0528 was built by DeepSeek. Chinese AI lab known for efficient, open-weight models. Gained attention for strong performance at lower cost. **Q: What benchmark scores did DeepSeek R1-0528 get?** A: DeepSeek R1-0528 reports 1 tracked benchmark score — BullshitBench v2: 8%. Scores are the figures published at release by DeepSeek. **Q: Is DeepSeek R1-0528 open source?** A: Partly. DeepSeek R1-0528 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after DeepSeek R1-0528?** A: DeepSeek's previous tracked release was DeepSeek V3-0324 on Mar 24 2025, 65 days earlier. It was followed by DeepSeek V3.1 on Aug 21 2025. ### Claude Sonnet 4 - URL: https://aireleasetracker.com/model/anthropic/claude-sonnet-4 - Developer: Anthropic - Release date: Thursday, May 22 2025 - Licensing: Proprietary - Benchmarks published at release: - BullshitBench v2: 30% - SWE-Bench Verified: 72.7% - GPQA Diamond: 75.4% **Q: When was Claude Sonnet 4 released?** A: Claude Sonnet 4 was released by Anthropic on Thursday, May 22 2025. **Q: Who made Claude Sonnet 4?** A: Claude Sonnet 4 was built by Anthropic. AI safety company building the Claude family of models. Founded in 2021 by former OpenAI researchers. **Q: What benchmark scores did Claude Sonnet 4 get?** A: Claude Sonnet 4 reports 3 tracked benchmark scores — BullshitBench v2: 30%; SWE-Bench Verified: 72.7%; GPQA Diamond: 75.4%. Scores are the figures published at release by Anthropic. **Q: Is Claude Sonnet 4 open source?** A: No. Claude Sonnet 4 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Claude Sonnet 4?** A: Anthropic's previous tracked release was Claude 3.7 Sonnet on Feb 24 2025, 87 days earlier. It was followed by Claude Opus 4 on May 22 2025. ### Claude Opus 4 - URL: https://aireleasetracker.com/model/anthropic/claude-opus-4 - Developer: Anthropic - Release date: Thursday, May 22 2025 - Licensing: Proprietary - Benchmarks published at release: - BullshitBench v2: 34% - SWE-Bench Verified: 72.5% - GPQA Diamond: 79.6% Claude Opus 4 arrived May 22, 2025 together with Claude Sonnet 4, reviving the Opus tier after more than a year and marking Anthropic's pivot to long-horizon agentic work. It was pitched at sustained, multi-hour coding sessions — refactoring across large codebases, running in agent harnesses, and using tools in extended loops — rather than single-shot chat answers. It posted 79.6% on GPQA Diamond and 72.5% on SWE-Bench Verified, roughly ten points above Claude 3.7 Sonnet on software engineering. Claude Opus 4.1 followed on August 5, 2025 as an incremental refinement, and the Opus line has since been Anthropic's flagship tier through Opus 4.5 and beyond. **Q: When was Claude Opus 4 released?** A: Claude Opus 4 was released by Anthropic on Thursday, May 22 2025. **Q: Who made Claude Opus 4?** A: Claude Opus 4 was built by Anthropic. AI safety company building the Claude family of models. Founded in 2021 by former OpenAI researchers. **Q: What benchmark scores did Claude Opus 4 get?** A: Claude Opus 4 reports 3 tracked benchmark scores — BullshitBench v2: 34%; SWE-Bench Verified: 72.5%; GPQA Diamond: 79.6%. Scores are the figures published at release by Anthropic. **Q: Is Claude Opus 4 open source?** A: No. Claude Opus 4 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Claude Opus 4?** A: Anthropic's previous tracked release was Claude Sonnet 4 on May 22 2025. It was followed by Claude Opus 4.1 on Aug 5 2025. ### Devstral Small - URL: https://aireleasetracker.com/model/mistral/devstral-small - Developer: Mistral - Release date: Wednesday, May 21 2025 - Licensing: Open Weight - Parameters: 24B - Benchmarks: none published in this dataset **Q: When was Devstral Small released?** A: Devstral Small was released by Mistral on Wednesday, May 21 2025. **Q: Who made Devstral Small?** A: Devstral Small was built by Mistral. French AI company building open and commercial models. Founded in 2023 by former Meta and DeepMind researchers. **Q: How many parameters does Devstral Small have?** A: Devstral Small is reported at 24B parameters. **Q: Is Devstral Small open source?** A: Partly. Devstral Small is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Devstral Small?** A: Mistral's previous tracked release was Mistral Medium 3 on May 7 2025, 14 days earlier. It was followed by Magistral Small on Jun 10 2025. ### Mistral Medium 3 - URL: https://aireleasetracker.com/model/mistral/mistral-medium-3 - Developer: Mistral - Release date: Wednesday, May 7 2025 - Licensing: Proprietary - Benchmarks: none published in this dataset **Q: When was Mistral Medium 3 released?** A: Mistral Medium 3 was released by Mistral on Wednesday, May 7 2025. **Q: Who made Mistral Medium 3?** A: Mistral Medium 3 was built by Mistral. French AI company building open and commercial models. Founded in 2023 by former Meta and DeepMind researchers. **Q: Is Mistral Medium 3 open source?** A: No. Mistral Medium 3 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Mistral Medium 3?** A: Mistral's previous tracked release was Mistral Small 3.1 on Mar 17 2025, 51 days earlier. It was followed by Devstral Small on May 21 2025. ### Qwen3 - URL: https://aireleasetracker.com/model/qwen/qwen3 - Developer: Qwen - Release date: Tuesday, Apr 29 2025 - Licensing: Open Weight - Parameters: 235B - Context window: 128k - Benchmarks: none published in this dataset Qwen3, released April 29, 2025, unified two things Alibaba had previously shipped separately: a conventional instruct model and a reasoning model. Every Qwen3 model could switch between thinking and non-thinking mode within a single set of weights, with the thinking budget controllable per request — the same hybrid idea Anthropic had introduced with Claude 3.7 Sonnet two months earlier, but in open weights. The family spanned dense models at 0.6B, 1.7B, 4B, 8B, 14B and 32B plus two mixture-of-experts models, 30B-A3B and the 235B-A22B flagship, all trained on 36 trillion tokens across 119 languages and dialects. Crucially, all of it went out under Apache 2.0 including the flagship — the first time Qwen had put the largest model of a generation under a fully permissive licence, and a direct contrast with Meta's custom Llama terms. **Q: When was Qwen3 released?** A: Qwen3 was released by Qwen on Tuesday, Apr 29 2025. **Q: Who made Qwen3?** A: Qwen3 was built by Qwen. Alibaba's AI lab, building the Qwen family. The most prolific publisher of open-weight models of any major lab, alongside a proprietary Max and Plus tier sold through Alibaba Cloud. **Q: What is the context window of Qwen3?** A: Qwen3 has a context window of 128k. That is the maximum amount of input plus output the model can hold in a single request. **Q: How many parameters does Qwen3 have?** A: Qwen3 is reported at 235B parameters. **Q: Is Qwen3 open source?** A: Partly. Qwen3 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Qwen3?** A: Qwen's previous tracked release was QwQ-32B on Mar 6 2025, 54 days earlier. It was followed by Qwen3-Coder on Jul 22 2025. ### Gemini 2.5 Flash - URL: https://aireleasetracker.com/model/google/gemini-2.5-flash - Developer: Google - Release date: Thursday, Apr 17 2025 - Licensing: Proprietary - Benchmarks published at release: - BullshitBench v2: 19% Gemini 2.5 Flash, released April 17, 2025, brought hybrid reasoning to Google's high-volume tier: developers could turn thinking on or off and set an exact token budget for it, trading accuracy against latency and cost per request. It was the first model to expose that dial explicitly, and the design was widely copied. Positioned as the workhorse of the 2.5 generation, Flash delivered a large share of 2.5 Pro's capability at a fraction of the price, making it one of the most heavily used API models of 2025 for chat products, summarisation, and agent sub-tasks. Flash-Lite followed in June 2025 to cover the ultra-low-cost end of the lineup. **Q: When was Gemini 2.5 Flash released?** A: Gemini 2.5 Flash was released by Google on Thursday, Apr 17 2025. **Q: Who made Gemini 2.5 Flash?** A: Gemini 2.5 Flash was built by Google. Builds the Gemini family of models through Google DeepMind. Integrates AI across Google products. **Q: What benchmark scores did Gemini 2.5 Flash get?** A: Gemini 2.5 Flash reports 1 tracked benchmark score — BullshitBench v2: 19%. Scores are the figures published at release by Google. **Q: Is Gemini 2.5 Flash open source?** A: No. Gemini 2.5 Flash is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Gemini 2.5 Flash?** A: Google's previous tracked release was Gemini 2.5 Pro on Mar 25 2025, 23 days earlier. It was followed by Gemini 2.5 Flash-Lite on Jun 17 2025. ### o3 - URL: https://aireleasetracker.com/model/openai/o3 - Developer: OpenAI - Release date: Wednesday, Apr 16 2025 - Licensing: Proprietary - Benchmarks published at release: - BullshitBench v2: 26% o3, released April 16, 2025 alongside o4-mini, was OpenAI's most capable reasoning model of the o-series era and the first to combine extended reasoning with full agentic tool use in ChatGPT — searching the web, running Python, and analysing images inside its chain of thought. First previewed in December 2024 with breakthrough ARC-AGI results, it took four months to reach general availability. In practice o3 became the model of choice for hard analytical work — mathematics, research questions, multi-step debugging — while GPT-4.1 handled everyday tasks. o3-pro followed in June 2025 for even longer reasoning budgets. The separate o-series branding ended when GPT-5 unified fast and reasoning models under one system in August 2025. **Q: When was o3 released?** A: o3 was released by OpenAI on Wednesday, Apr 16 2025. **Q: Who made o3?** A: o3 was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: What benchmark scores did o3 get?** A: o3 reports 1 tracked benchmark score — BullshitBench v2: 26%. Scores are the figures published at release by OpenAI. **Q: Is o3 open source?** A: No. o3 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after o3?** A: OpenAI's previous tracked release was GPT-4.1 nano on Apr 14 2025, 2 days earlier. It was followed by o4-mini on Apr 16 2025. ### o4-mini - URL: https://aireleasetracker.com/model/openai/o4-mini - Developer: OpenAI - Release date: Wednesday, Apr 16 2025 - Licensing: Proprietary - Benchmarks published at release: - BullshitBench v2: 8% **Q: When was o4-mini released?** A: o4-mini was released by OpenAI on Wednesday, Apr 16 2025. **Q: Who made o4-mini?** A: o4-mini was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: What benchmark scores did o4-mini get?** A: o4-mini reports 1 tracked benchmark score — BullshitBench v2: 8%. Scores are the figures published at release by OpenAI. **Q: Is o4-mini open source?** A: No. o4-mini is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after o4-mini?** A: OpenAI's previous tracked release was o3 on Apr 16 2025. It was followed by o4-mini-high on Apr 16 2025. ### o4-mini-high - URL: https://aireleasetracker.com/model/openai/o4-mini-high - Developer: OpenAI - Release date: Wednesday, Apr 16 2025 - Licensing: Proprietary - Benchmarks: none published in this dataset **Q: When was o4-mini-high released?** A: o4-mini-high was released by OpenAI on Wednesday, Apr 16 2025. **Q: Who made o4-mini-high?** A: o4-mini-high was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: Is o4-mini-high open source?** A: No. o4-mini-high is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after o4-mini-high?** A: OpenAI's previous tracked release was o4-mini on Apr 16 2025. It was followed by o3-pro on Jun 10 2025. ### GPT-4.1 - URL: https://aireleasetracker.com/model/openai/gpt-4.1 - Developer: OpenAI - Release date: Monday, Apr 14 2025 - Licensing: Proprietary - Benchmarks published at release: - BullshitBench v2: 14% **Q: When was GPT-4.1 released?** A: GPT-4.1 was released by OpenAI on Monday, Apr 14 2025. **Q: Who made GPT-4.1?** A: GPT-4.1 was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: What benchmark scores did GPT-4.1 get?** A: GPT-4.1 reports 1 tracked benchmark score — BullshitBench v2: 14%. Scores are the figures published at release by OpenAI. **Q: Is GPT-4.1 open source?** A: No. GPT-4.1 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after GPT-4.1?** A: OpenAI's previous tracked release was o1-pro on Mar 19 2025, 26 days earlier. It was followed by GPT-4.1 mini on Apr 14 2025. ### GPT-4.1 mini - URL: https://aireleasetracker.com/model/openai/gpt-4.1-mini - Developer: OpenAI - Release date: Monday, Apr 14 2025 - Licensing: Proprietary - Benchmarks: none published in this dataset **Q: When was GPT-4.1 mini released?** A: GPT-4.1 mini was released by OpenAI on Monday, Apr 14 2025. **Q: Who made GPT-4.1 mini?** A: GPT-4.1 mini was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: Is GPT-4.1 mini open source?** A: No. GPT-4.1 mini is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after GPT-4.1 mini?** A: OpenAI's previous tracked release was GPT-4.1 on Apr 14 2025. It was followed by GPT-4.1 nano on Apr 14 2025. ### GPT-4.1 nano - URL: https://aireleasetracker.com/model/openai/gpt-4.1-nano - Developer: OpenAI - Release date: Monday, Apr 14 2025 - Licensing: Proprietary - Context window: 1M - Benchmarks: none published in this dataset **Q: When was GPT-4.1 nano released?** A: GPT-4.1 nano was released by OpenAI on Monday, Apr 14 2025. **Q: Who made GPT-4.1 nano?** A: GPT-4.1 nano was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: What is the context window of GPT-4.1 nano?** A: GPT-4.1 nano has a context window of 1M. That is the maximum amount of input plus output the model can hold in a single request. **Q: Is GPT-4.1 nano open source?** A: No. GPT-4.1 nano is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after GPT-4.1 nano?** A: OpenAI's previous tracked release was GPT-4.1 mini on Apr 14 2025. It was followed by o3 on Apr 16 2025. ### Kimi-VL - URL: https://aireleasetracker.com/model/moonshot/kimi-vl - Developer: Moonshot AI - Release date: Thursday, Apr 10 2025 - Licensing: Open Weight - Parameters: 16B - Benchmarks: none published in this dataset **Q: When was Kimi-VL released?** A: Kimi-VL was released by Moonshot AI on Thursday, Apr 10 2025. **Q: Who made Kimi-VL?** A: Kimi-VL was built by Moonshot AI. Chinese AI company building the Kimi family of models. Founded in 2023, known for long-context and efficient MoE architectures. **Q: How many parameters does Kimi-VL have?** A: Kimi-VL is reported at 16B parameters. **Q: Is Kimi-VL open source?** A: Partly. Kimi-VL is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Kimi-VL?** A: Moonshot AI's previous tracked release was Kimi K1.5 on Jan 20 2025, 80 days earlier. It was followed by Kimi-Dev on Jun 17 2025. ### LLaMA 4 Scout - URL: https://aireleasetracker.com/model/meta/llama-4-scout - Developer: Meta - Release date: Saturday, Apr 5 2025 - Licensing: Open Weight - Benchmarks published at release: - BullshitBench v2: 19% **Q: When was LLaMA 4 Scout released?** A: LLaMA 4 Scout was released by Meta on Saturday, Apr 5 2025. **Q: Who made LLaMA 4 Scout?** A: LLaMA 4 Scout was built by Meta. Develops the open-weight Llama series of models. Committed to open-source AI research. **Q: What benchmark scores did LLaMA 4 Scout get?** A: LLaMA 4 Scout reports 1 tracked benchmark score — BullshitBench v2: 19%. Scores are the figures published at release by Meta. **Q: Is LLaMA 4 Scout open source?** A: Partly. LLaMA 4 Scout is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after LLaMA 4 Scout?** A: Meta's previous tracked release was LLaMA 3.3 on Dec 6 2024, 120 days earlier. It was followed by LLaMA 4 Maverick on Apr 5 2025. ### LLaMA 4 Maverick - URL: https://aireleasetracker.com/model/meta/llama-4-maverick - Developer: Meta - Release date: Saturday, Apr 5 2025 - Licensing: Open Weight - Benchmarks published at release: - BullshitBench v2: 28% Llama 4 Maverick, released April 5, 2025 alongside the smaller Llama 4 Scout, moved the Llama line to a natively multimodal mixture-of-experts design — roughly 400B total parameters with 17B active per token. It launched with claims of beating GPT-4o and Gemini 2.0 Flash at far lower serving cost. The release proved controversial: an experimental chat-tuned variant's LMArena ranking drew accusations of benchmark gaming, and the promised frontier-scale Behemoth model never shipped. Llama 4 ended up the final release of the Llama era — Meta's next generation arrived in April 2026 under the new Muse Spark name. **Q: When was LLaMA 4 Maverick released?** A: LLaMA 4 Maverick was released by Meta on Saturday, Apr 5 2025. **Q: Who made LLaMA 4 Maverick?** A: LLaMA 4 Maverick was built by Meta. Develops the open-weight Llama series of models. Committed to open-source AI research. **Q: What benchmark scores did LLaMA 4 Maverick get?** A: LLaMA 4 Maverick reports 1 tracked benchmark score — BullshitBench v2: 28%. Scores are the figures published at release by Meta. **Q: Is LLaMA 4 Maverick open source?** A: Partly. LLaMA 4 Maverick is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after LLaMA 4 Maverick?** A: Meta's previous tracked release was LLaMA 4 Scout on Apr 5 2025. It was followed by Muse Spark on Apr 8 2026. ### Gemini 2.5 Pro - URL: https://aireleasetracker.com/model/google/gemini-2.5-pro - Developer: Google - Release date: Tuesday, Mar 25 2025 - Licensing: Proprietary - Benchmarks published at release: - BullshitBench v2: 20% - SWE-Bench Verified: 59.6% - GPQA Diamond: 86.4% - MMMU: 68% - Arena Elo (Code): 1224 Gemini 2.5 Pro, released March 25, 2025, was Google's first thinking-native flagship — reasoning built into the model rather than offered as a separate mode — and it debuted at the top of the LMArena leaderboard by a wide margin. It combined a 1M-token context window with 86.4% on GPQA Diamond, at the time the strongest published science-reasoning score of any model. It was arguably Google's definitive 2025 comeback: after a year of trailing OpenAI and Anthropic in developer mindshare, 2.5 Pro made Gemini a serious default for coding, long-document analysis, and video understanding. It held Google's flagship position for nearly eight months until Gemini 3.0 Pro arrived in November 2025. **Q: When was Gemini 2.5 Pro released?** A: Gemini 2.5 Pro was released by Google on Tuesday, Mar 25 2025. **Q: Who made Gemini 2.5 Pro?** A: Gemini 2.5 Pro was built by Google. Builds the Gemini family of models through Google DeepMind. Integrates AI across Google products. **Q: What benchmark scores did Gemini 2.5 Pro get?** A: Gemini 2.5 Pro reports 5 tracked benchmark scores — BullshitBench v2: 20%; SWE-Bench Verified: 59.6%; GPQA Diamond: 86.4%; MMMU: 68%; Arena Elo (Code): 1224. Scores are the figures published at release by Google. **Q: Is Gemini 2.5 Pro open source?** A: No. Gemini 2.5 Pro is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Gemini 2.5 Pro?** A: Google's previous tracked release was Gemma 3 on Mar 12 2025, 13 days earlier. It was followed by Gemini 2.5 Flash on Apr 17 2025. ### DeepSeek V3-0324 - URL: https://aireleasetracker.com/model/deepseek/deepseek-v3-0324 - Developer: DeepSeek - Release date: Monday, Mar 24 2025 - Licensing: Open Weight - Benchmarks: none published in this dataset **Q: When was DeepSeek V3-0324 released?** A: DeepSeek V3-0324 was released by DeepSeek on Monday, Mar 24 2025. **Q: Who made DeepSeek V3-0324?** A: DeepSeek V3-0324 was built by DeepSeek. Chinese AI lab known for efficient, open-weight models. Gained attention for strong performance at lower cost. **Q: Is DeepSeek V3-0324 open source?** A: Partly. DeepSeek V3-0324 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after DeepSeek V3-0324?** A: DeepSeek's previous tracked release was DeepSeek Chat (R1-based) on Jan 20 2025, 63 days earlier. It was followed by DeepSeek R1-0528 on May 28 2025. ### o1-pro - URL: https://aireleasetracker.com/model/openai/o1-pro - Developer: OpenAI - Release date: Wednesday, Mar 19 2025 - Licensing: Proprietary - Benchmarks: none published in this dataset **Q: When was o1-pro released?** A: o1-pro was released by OpenAI on Wednesday, Mar 19 2025. **Q: Who made o1-pro?** A: o1-pro was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: Is o1-pro open source?** A: No. o1-pro is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after o1-pro?** A: OpenAI's previous tracked release was GPT-4.5 on Feb 27 2025, 20 days earlier. It was followed by GPT-4.1 on Apr 14 2025. ### Mistral Small 3.1 - URL: https://aireleasetracker.com/model/mistral/mistral-small-3.1 - Developer: Mistral - Release date: Monday, Mar 17 2025 - Licensing: Open Weight - Benchmarks: none published in this dataset **Q: When was Mistral Small 3.1 released?** A: Mistral Small 3.1 was released by Mistral on Monday, Mar 17 2025. **Q: Who made Mistral Small 3.1?** A: Mistral Small 3.1 was built by Mistral. French AI company building open and commercial models. Founded in 2023 by former Meta and DeepMind researchers. **Q: Is Mistral Small 3.1 open source?** A: Partly. Mistral Small 3.1 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Mistral Small 3.1?** A: Mistral's previous tracked release was Mistral Small 3 on Jan 30 2025, 46 days earlier. It was followed by Mistral Medium 3 on May 7 2025. ### Gemma 3 - URL: https://aireleasetracker.com/model/google/gemma-3 - Developer: Google - Release date: Wednesday, Mar 12 2025 - Licensing: Open Weight - Benchmarks: none published in this dataset Gemma 3, released March 12, 2025, is Google's open-weight family distilled from Gemini technology, spanning roughly 1B to 27B parameters with vision support and a 128K context window on the larger sizes. Its design target was practical: state-of-the-art quality for a model that runs on a single GPU or TPU. The 27B variant punched far above its weight on human-preference leaderboards, rivalling much larger open models, and the family's permissive terms and small-hardware footprint made it a staple for on-device and self-hosted deployments. Gemma 4 continued the line in April 2026. **Q: When was Gemma 3 released?** A: Gemma 3 was released by Google on Wednesday, Mar 12 2025. **Q: Who made Gemma 3?** A: Gemma 3 was built by Google. Builds the Gemini family of models through Google DeepMind. Integrates AI across Google products. **Q: Is Gemma 3 open source?** A: Partly. Gemma 3 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Gemma 3?** A: Google's previous tracked release was Gemini 2.0 Pro on Feb 5 2025, 35 days earlier. It was followed by Gemini 2.5 Pro on Mar 25 2025. ### QwQ-32B - URL: https://aireleasetracker.com/model/qwen/qwq-32b - Developer: Qwen - Release date: Thursday, Mar 6 2025 - Licensing: Open Weight - Parameters: 32B - Context window: 128k - Benchmarks: none published in this dataset QwQ-32B, released March 6, 2025 under Apache 2.0, was Qwen's answer to the reasoning-model wave that DeepSeek-R1 had set off six weeks earlier. It followed QwQ-32B-Preview from November 2024, and where the preview was explicitly experimental, the March release was a finished reinforcement-learning-trained reasoning model with a 128K context window. The pitch was compression rather than scale: Alibaba positioned a 32B dense model as competitive with DeepSeek-R1, a 671B mixture-of-experts model, on mathematics and coding evaluations. That mattered practically as much as competitively — a 32B model runs on a single high-end GPU, so it put a frontier-style reasoning model within reach of individuals rather than clusters. The approach was folded into the main line two months later when Qwen3 shipped with switchable thinking modes built in. **Q: When was QwQ-32B released?** A: QwQ-32B was released by Qwen on Thursday, Mar 6 2025. **Q: Who made QwQ-32B?** A: QwQ-32B was built by Qwen. Alibaba's AI lab, building the Qwen family. The most prolific publisher of open-weight models of any major lab, alongside a proprietary Max and Plus tier sold through Alibaba Cloud. **Q: What is the context window of QwQ-32B?** A: QwQ-32B has a context window of 128k. That is the maximum amount of input plus output the model can hold in a single request. **Q: How many parameters does QwQ-32B have?** A: QwQ-32B is reported at 32B parameters. **Q: Is QwQ-32B open source?** A: Partly. QwQ-32B is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after QwQ-32B?** A: Qwen's previous tracked release was Qwen2.5-Max on Jan 28 2025, 37 days earlier. It was followed by Qwen3 on Apr 29 2025. ### GPT-4.5 - URL: https://aireleasetracker.com/model/openai/gpt-4.5 - Developer: OpenAI - Release date: Thursday, Feb 27 2025 - Licensing: Proprietary - Benchmarks: none published in this dataset **Q: When was GPT-4.5 released?** A: GPT-4.5 was released by OpenAI on Thursday, Feb 27 2025. **Q: Who made GPT-4.5?** A: GPT-4.5 was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: Is GPT-4.5 open source?** A: No. GPT-4.5 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after GPT-4.5?** A: OpenAI's previous tracked release was o3-mini-high on Jan 31 2025, 27 days earlier. It was followed by o1-pro on Mar 19 2025. ### Claude 3.7 Sonnet - URL: https://aireleasetracker.com/model/anthropic/claude-3.7-sonnet - Developer: Anthropic - Release date: Monday, Feb 24 2025 - Licensing: Proprietary - Benchmarks published at release: - BullshitBench v2: 49% - SWE-Bench Verified: 62.3% - GPQA Diamond: 68% Claude 3.7 Sonnet, released February 24, 2025, was Anthropic's first hybrid reasoning model: a single model that could answer instantly or engage an extended-thinking mode where it reasons step by step before responding, with the thinking budget under API control. It launched alongside the first research preview of Claude Code, Anthropic's agentic command-line coding tool. The reasoning upgrade showed up most clearly in software engineering, where Claude 3.7 Sonnet scored 62.3% on SWE-Bench Verified — up from 49.0% for its predecessor and the best published score of any model at the time. GPQA Diamond rose to 68.0%. It held the top of the Claude lineup for three months until Claude Sonnet 4 and Opus 4 arrived in May 2025. **Q: When was Claude 3.7 Sonnet released?** A: Claude 3.7 Sonnet was released by Anthropic on Monday, Feb 24 2025. **Q: Who made Claude 3.7 Sonnet?** A: Claude 3.7 Sonnet was built by Anthropic. AI safety company building the Claude family of models. Founded in 2021 by former OpenAI researchers. **Q: What benchmark scores did Claude 3.7 Sonnet get?** A: Claude 3.7 Sonnet reports 3 tracked benchmark scores — BullshitBench v2: 49%; SWE-Bench Verified: 62.3%; GPQA Diamond: 68%. Scores are the figures published at release by Anthropic. **Q: Is Claude 3.7 Sonnet open source?** A: No. Claude 3.7 Sonnet is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Claude 3.7 Sonnet?** A: Anthropic's previous tracked release was Claude 3.5 Sonnet (upgraded) on Oct 22 2024, 125 days earlier. It was followed by Claude Sonnet 4 on May 22 2025. ### Grok 3 - URL: https://aireleasetracker.com/model/xai/grok-3 - Developer: SpaceXAI - Release date: Monday, Feb 17 2025 - Licensing: Proprietary - Benchmarks: none published in this dataset Grok 3, released February 17, 2025 alongside Grok 3 mini, was trained on Colossus — xAI's Memphis supercomputer of around 200,000 GPUs — with roughly ten times the compute of Grok 2. It introduced Think mode for step-by-step reasoning and DeepSearch, an agentic research tool that scans the web and X in real time. The release put xAI on the frontier map barely two years after the company's founding, with launch benchmarks competitive with o1 and Gemini 2.0 across math, science, and coding. It headlined xAI's lineup for five months until Grok 4 arrived in July 2025. **Q: When was Grok 3 released?** A: Grok 3 was released by SpaceXAI on Monday, Feb 17 2025. **Q: Who made Grok 3?** A: Grok 3 was built by SpaceXAI. Elon Musk's AI company building the Grok series of models. Founded in 2023 as xAI, now part of SpaceX. **Q: Is Grok 3 open source?** A: No. Grok 3 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Grok 3?** A: SpaceXAI's previous tracked release was Grok‑2 on Aug 14 2024, 187 days earlier. It was followed by Grok 3 mini on Feb 17 2025. ### Grok 3 mini - URL: https://aireleasetracker.com/model/xai/grok-3-mini - Developer: SpaceXAI - Release date: Monday, Feb 17 2025 - Licensing: Proprietary - Benchmarks: none published in this dataset **Q: When was Grok 3 mini released?** A: Grok 3 mini was released by SpaceXAI on Monday, Feb 17 2025. **Q: Who made Grok 3 mini?** A: Grok 3 mini was built by SpaceXAI. Elon Musk's AI company building the Grok series of models. Founded in 2023 as xAI, now part of SpaceX. **Q: Is Grok 3 mini open source?** A: No. Grok 3 mini is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Grok 3 mini?** A: SpaceXAI's previous tracked release was Grok 3 on Feb 17 2025. It was followed by Grok 4 on Jul 9 2025. ### Gemini 2.0 Flash-Lite - URL: https://aireleasetracker.com/model/google/gemini-2.0-flash-lite - Developer: Google - Release date: Wednesday, Feb 5 2025 - Licensing: Proprietary - Benchmarks: none published in this dataset **Q: When was Gemini 2.0 Flash-Lite released?** A: Gemini 2.0 Flash-Lite was released by Google on Wednesday, Feb 5 2025. **Q: Who made Gemini 2.0 Flash-Lite?** A: Gemini 2.0 Flash-Lite was built by Google. Builds the Gemini family of models through Google DeepMind. Integrates AI across Google products. **Q: Is Gemini 2.0 Flash-Lite open source?** A: No. Gemini 2.0 Flash-Lite is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Gemini 2.0 Flash-Lite?** A: Google's previous tracked release was Gemini 2.0 Flash on Jan 30 2025, 6 days earlier. It was followed by Gemini 2.0 Pro on Feb 5 2025. ### Gemini 2.0 Pro - URL: https://aireleasetracker.com/model/google/gemini-2.0-pro - Developer: Google - Release date: Wednesday, Feb 5 2025 - Licensing: Proprietary - Benchmarks: none published in this dataset **Q: When was Gemini 2.0 Pro released?** A: Gemini 2.0 Pro was released by Google on Wednesday, Feb 5 2025. **Q: Who made Gemini 2.0 Pro?** A: Gemini 2.0 Pro was built by Google. Builds the Gemini family of models through Google DeepMind. Integrates AI across Google products. **Q: Is Gemini 2.0 Pro open source?** A: No. Gemini 2.0 Pro is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Gemini 2.0 Pro?** A: Google's previous tracked release was Gemini 2.0 Flash-Lite on Feb 5 2025. It was followed by Gemma 3 on Mar 12 2025. ### o3-mini - URL: https://aireleasetracker.com/model/openai/o3-mini - Developer: OpenAI - Release date: Friday, Jan 31 2025 - Licensing: Proprietary - Benchmarks: none published in this dataset **Q: When was o3-mini released?** A: o3-mini was released by OpenAI on Friday, Jan 31 2025. **Q: Who made o3-mini?** A: o3-mini was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: Is o3-mini open source?** A: No. o3-mini is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after o3-mini?** A: OpenAI's previous tracked release was o1 on Dec 5 2024, 57 days earlier. It was followed by o3-mini-high on Jan 31 2025. ### o3-mini-high - URL: https://aireleasetracker.com/model/openai/o3-mini-high - Developer: OpenAI - Release date: Friday, Jan 31 2025 - Licensing: Proprietary - Benchmarks: none published in this dataset **Q: When was o3-mini-high released?** A: o3-mini-high was released by OpenAI on Friday, Jan 31 2025. **Q: Who made o3-mini-high?** A: o3-mini-high was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: Is o3-mini-high open source?** A: No. o3-mini-high is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after o3-mini-high?** A: OpenAI's previous tracked release was o3-mini on Jan 31 2025. It was followed by GPT-4.5 on Feb 27 2025. ### Gemini 2.0 Flash - URL: https://aireleasetracker.com/model/google/gemini-2.0-flash - Developer: Google - Release date: Thursday, Jan 30 2025 - Licensing: Proprietary - Benchmarks published at release: - BullshitBench v2: 15% Gemini 2.0 Flash, generally available January 30, 2025 after a December 2024 preview, was the launch model of what Google called the agentic era: native tool use, multimodal output including image generation and steerable text-to-speech, and a Multimodal Live API for real-time audio and video streaming — all at Flash-tier latency and pricing. Unusually, Google led the 2.0 generation with its workhorse model rather than its flagship; 2.0 Flash outperformed the previous-generation 1.5 Pro while costing far less, and it powered the first wave of Google's agent prototypes. The 2.0 line filled out with Flash-Lite and Pro within a week, but was quickly eclipsed by the thinking-native Gemini 2.5 family in March 2025. **Q: When was Gemini 2.0 Flash released?** A: Gemini 2.0 Flash was released by Google on Thursday, Jan 30 2025. **Q: Who made Gemini 2.0 Flash?** A: Gemini 2.0 Flash was built by Google. Builds the Gemini family of models through Google DeepMind. Integrates AI across Google products. **Q: What benchmark scores did Gemini 2.0 Flash get?** A: Gemini 2.0 Flash reports 1 tracked benchmark score — BullshitBench v2: 15%. Scores are the figures published at release by Google. **Q: Is Gemini 2.0 Flash open source?** A: No. Gemini 2.0 Flash is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Gemini 2.0 Flash?** A: Google's previous tracked release was Gemini 1.5 Flash-8B on Oct 3 2024, 119 days earlier. It was followed by Gemini 2.0 Flash-Lite on Feb 5 2025. ### Mistral Small 3 - URL: https://aireleasetracker.com/model/mistral/mistral-small-3 - Developer: Mistral - Release date: Thursday, Jan 30 2025 - Licensing: Open Weight - Benchmarks: none published in this dataset **Q: When was Mistral Small 3 released?** A: Mistral Small 3 was released by Mistral on Thursday, Jan 30 2025. **Q: Who made Mistral Small 3?** A: Mistral Small 3 was built by Mistral. French AI company building open and commercial models. Founded in 2023 by former Meta and DeepMind researchers. **Q: Is Mistral Small 3 open source?** A: Partly. Mistral Small 3 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Mistral Small 3?** A: Mistral's previous tracked release was Codestral 25.01 on Jan 13 2025, 17 days earlier. It was followed by Mistral Small 3.1 on Mar 17 2025. ### Qwen2.5-Max - URL: https://aireleasetracker.com/model/qwen/qwen2.5-max - Developer: Qwen - Release date: Tuesday, Jan 28 2025 - Licensing: Proprietary - Context window: 32k - Benchmarks: none published in this dataset **Q: When was Qwen2.5-Max released?** A: Qwen2.5-Max was released by Qwen on Tuesday, Jan 28 2025. **Q: Who made Qwen2.5-Max?** A: Qwen2.5-Max was built by Qwen. Alibaba's AI lab, building the Qwen family. The most prolific publisher of open-weight models of any major lab, alongside a proprietary Max and Plus tier sold through Alibaba Cloud. **Q: What is the context window of Qwen2.5-Max?** A: Qwen2.5-Max has a context window of 32k. That is the maximum amount of input plus output the model can hold in a single request. **Q: Is Qwen2.5-Max open source?** A: No. Qwen2.5-Max is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Qwen2.5-Max?** A: Qwen's previous tracked release was QwQ-32B-Preview on Nov 28 2024, 61 days earlier. It was followed by QwQ-32B on Mar 6 2025. ### DeepSeek Chat (R1-based) - URL: https://aireleasetracker.com/model/deepseek/deepseek-chat-(r1-based) - Developer: DeepSeek - Release date: Monday, Jan 20 2025 - Licensing: Open Weight - Benchmarks: none published in this dataset DeepSeek R1, released January 20, 2025, matched OpenAI's o1 on reasoning benchmarks with fully open weights under an MIT license — and, unlike o1, exposed its complete chain of thought. Its training recipe was the real breakthrough: reasoning learned primarily through reinforcement learning on the V3 base, with the R1-Zero experiment showing reasoning could emerge from RL alone. The release caused the single largest one-day market reaction in AI history, wiping hundreds of billions from US tech valuations as investors reassessed what frontier capability costs. Its distilled variants brought reasoning to consumer hardware, and the updated R1-0528 followed in May 2025. **Q: When was DeepSeek Chat (R1-based) released?** A: DeepSeek Chat (R1-based) was released by DeepSeek on Monday, Jan 20 2025. **Q: Who made DeepSeek Chat (R1-based)?** A: DeepSeek Chat (R1-based) was built by DeepSeek. Chinese AI lab known for efficient, open-weight models. Gained attention for strong performance at lower cost. **Q: Is DeepSeek Chat (R1-based) open source?** A: Partly. DeepSeek Chat (R1-based) is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after DeepSeek Chat (R1-based)?** A: DeepSeek's previous tracked release was DeepSeek V3 Chat on Dec 26 2024, 25 days earlier. It was followed by DeepSeek V3-0324 on Mar 24 2025. ### Kimi K1.5 - URL: https://aireleasetracker.com/model/moonshot/kimi-k1.5 - Developer: Moonshot AI - Release date: Monday, Jan 20 2025 - Licensing: Proprietary - Context window: 128k - Benchmarks: none published in this dataset **Q: When was Kimi K1.5 released?** A: Kimi K1.5 was released by Moonshot AI on Monday, Jan 20 2025. **Q: Who made Kimi K1.5?** A: Kimi K1.5 was built by Moonshot AI. Chinese AI company building the Kimi family of models. Founded in 2023, known for long-context and efficient MoE architectures. **Q: What is the context window of Kimi K1.5?** A: Kimi K1.5 has a context window of 128k. That is the maximum amount of input plus output the model can hold in a single request. **Q: Is Kimi K1.5 open source?** A: No. Kimi K1.5 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Kimi K1.5?** A: Moonshot AI's previous tracked release was Kimi Chat on Oct 9 2023, 469 days earlier. It was followed by Kimi-VL on Apr 10 2025. ### Codestral 25.01 - URL: https://aireleasetracker.com/model/mistral/codestral-25.01 - Developer: Mistral - Release date: Monday, Jan 13 2025 - Licensing: Proprietary - Context window: 256k - Benchmarks: none published in this dataset **Q: When was Codestral 25.01 released?** A: Codestral 25.01 was released by Mistral on Monday, Jan 13 2025. **Q: Who made Codestral 25.01?** A: Codestral 25.01 was built by Mistral. French AI company building open and commercial models. Founded in 2023 by former Meta and DeepMind researchers. **Q: What is the context window of Codestral 25.01?** A: Codestral 25.01 has a context window of 256k. That is the maximum amount of input plus output the model can hold in a single request. **Q: Is Codestral 25.01 open source?** A: No. Codestral 25.01 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Codestral 25.01?** A: Mistral's previous tracked release was Pixtral Large 24.11 on Nov 18 2024, 56 days earlier. It was followed by Mistral Small 3 on Jan 30 2025. ### DeepSeek V3 Chat - URL: https://aireleasetracker.com/model/deepseek/deepseek-v3-chat - Developer: DeepSeek - Release date: Thursday, Dec 26 2024 - Licensing: Open Weight - Benchmarks: none published in this dataset DeepSeek V3, released in December 2024, was the efficiency shock of its generation: a 671B-parameter mixture-of-experts model with 37B active per token, trained for a reported ~$5.6M in compute — one to two orders of magnitude less than Western frontier budgets — yet competitive with GPT-4o and Claude 3.5 Sonnet on standard benchmarks. Released as open weights under a permissive license, V3 introduced multi-head latent attention and aggressive FP8 training to the open ecosystem and became the base model for DeepSeek R1 a month later. Together the pair triggered a global repricing of AI training costs in early 2025. **Q: When was DeepSeek V3 Chat released?** A: DeepSeek V3 Chat was released by DeepSeek on Thursday, Dec 26 2024. **Q: Who made DeepSeek V3 Chat?** A: DeepSeek V3 Chat was built by DeepSeek. Chinese AI lab known for efficient, open-weight models. Gained attention for strong performance at lower cost. **Q: Is DeepSeek V3 Chat open source?** A: Partly. DeepSeek V3 Chat is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after DeepSeek V3 Chat?** A: DeepSeek's previous tracked release was DeepSeek V3 Base on Dec 25 2024, 1 day earlier. It was followed by DeepSeek Chat (R1-based) on Jan 20 2025. ### DeepSeek V3 Base - URL: https://aireleasetracker.com/model/deepseek/deepseek-v3-base - Developer: DeepSeek - Release date: Wednesday, Dec 25 2024 - Licensing: Open Weight - Benchmarks: none published in this dataset **Q: When was DeepSeek V3 Base released?** A: DeepSeek V3 Base was released by DeepSeek on Wednesday, Dec 25 2024. **Q: Who made DeepSeek V3 Base?** A: DeepSeek V3 Base was built by DeepSeek. Chinese AI lab known for efficient, open-weight models. Gained attention for strong performance at lower cost. **Q: Is DeepSeek V3 Base open source?** A: Partly. DeepSeek V3 Base is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after DeepSeek V3 Base?** A: DeepSeek's previous tracked release was DeepSeek V2.5 (Revised) on Dec 10 2024, 15 days earlier. It was followed by DeepSeek V3 Chat on Dec 26 2024. ### DeepSeek V2.5 (Revised) - URL: https://aireleasetracker.com/model/deepseek/deepseek-v2.5-(revised) - Developer: DeepSeek - Release date: Tuesday, Dec 10 2024 - Licensing: Open Weight - Benchmarks: none published in this dataset **Q: When was DeepSeek V2.5 (Revised) released?** A: DeepSeek V2.5 (Revised) was released by DeepSeek on Tuesday, Dec 10 2024. **Q: Who made DeepSeek V2.5 (Revised)?** A: DeepSeek V2.5 (Revised) was built by DeepSeek. Chinese AI lab known for efficient, open-weight models. Gained attention for strong performance at lower cost. **Q: Is DeepSeek V2.5 (Revised) open source?** A: Partly. DeepSeek V2.5 (Revised) is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after DeepSeek V2.5 (Revised)?** A: DeepSeek's previous tracked release was DeepSeek-R1-Lite Preview on Nov 20 2024, 20 days earlier. It was followed by DeepSeek V3 Base on Dec 25 2024. ### LLaMA 3.3 - URL: https://aireleasetracker.com/model/meta/llama-3.3 - Developer: Meta - Release date: Friday, Dec 6 2024 - Licensing: Open Weight - Parameters: 70B - Benchmarks: none published in this dataset **Q: When was LLaMA 3.3 released?** A: LLaMA 3.3 was released by Meta on Friday, Dec 6 2024. **Q: Who made LLaMA 3.3?** A: LLaMA 3.3 was built by Meta. Develops the open-weight Llama series of models. Committed to open-source AI research. **Q: How many parameters does LLaMA 3.3 have?** A: LLaMA 3.3 is reported at 70B parameters. **Q: Is LLaMA 3.3 open source?** A: Partly. LLaMA 3.3 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after LLaMA 3.3?** A: Meta's previous tracked release was LLaMA 3.2 on Sep 25 2024, 72 days earlier. It was followed by LLaMA 4 Scout on Apr 5 2025. ### o1 - URL: https://aireleasetracker.com/model/openai/o1 - Developer: OpenAI - Release date: Thursday, Dec 5 2024 - Licensing: Proprietary - Context window: 200k - Benchmarks published at release: - GPQA Diamond: 75.7% o1, released December 5, 2024 after a September preview, was the first production reasoning model: instead of answering immediately, it generates a long private chain of thought trained with reinforcement learning before responding. The approach produced dramatic gains on competition mathematics, science, and hard coding problems that had resisted ordinary scaling. o1 established the test-time-compute paradigm that reshaped the entire industry — within months every major lab shipped a reasoning model, from DeepSeek R1 to Gemini 2.5 Pro to Claude 3.7 Sonnet's extended thinking. Its own line moved fast too: o3-mini arrived in January 2025 and the full o3 in April 2025. **Q: When was o1 released?** A: o1 was released by OpenAI on Thursday, Dec 5 2024. **Q: Who made o1?** A: o1 was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: What benchmark scores did o1 get?** A: o1 reports 1 tracked benchmark score — GPQA Diamond: 75.7%. Scores are the figures published at release by OpenAI. **Q: What is the context window of o1?** A: o1 has a context window of 200k. That is the maximum amount of input plus output the model can hold in a single request. **Q: Is o1 open source?** A: No. o1 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after o1?** A: OpenAI's previous tracked release was o1-mini on Sep 12 2024, 84 days earlier. It was followed by o3-mini on Jan 31 2025. ### QwQ-32B-Preview - URL: https://aireleasetracker.com/model/qwen/qwq-32b-preview - Developer: Qwen - Release date: Thursday, Nov 28 2024 - Licensing: Open Weight - Parameters: 32B - Context window: 32k - Benchmarks: none published in this dataset **Q: When was QwQ-32B-Preview released?** A: QwQ-32B-Preview was released by Qwen on Thursday, Nov 28 2024. **Q: Who made QwQ-32B-Preview?** A: QwQ-32B-Preview was built by Qwen. Alibaba's AI lab, building the Qwen family. The most prolific publisher of open-weight models of any major lab, alongside a proprietary Max and Plus tier sold through Alibaba Cloud. **Q: What is the context window of QwQ-32B-Preview?** A: QwQ-32B-Preview has a context window of 32k. That is the maximum amount of input plus output the model can hold in a single request. **Q: How many parameters does QwQ-32B-Preview have?** A: QwQ-32B-Preview is reported at 32B parameters. **Q: Is QwQ-32B-Preview open source?** A: Partly. QwQ-32B-Preview is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after QwQ-32B-Preview?** A: Qwen's previous tracked release was Qwen2.5-Coder on Nov 12 2024, 16 days earlier. It was followed by Qwen2.5-Max on Jan 28 2025. ### DeepSeek-R1-Lite Preview - URL: https://aireleasetracker.com/model/deepseek/deepseek-r1-lite-preview - Developer: DeepSeek - Release date: Wednesday, Nov 20 2024 - Licensing: Open Weight - Benchmarks: none published in this dataset **Q: When was DeepSeek-R1-Lite Preview released?** A: DeepSeek-R1-Lite Preview was released by DeepSeek on Wednesday, Nov 20 2024. **Q: Who made DeepSeek-R1-Lite Preview?** A: DeepSeek-R1-Lite Preview was built by DeepSeek. Chinese AI lab known for efficient, open-weight models. Gained attention for strong performance at lower cost. **Q: Is DeepSeek-R1-Lite Preview open source?** A: Partly. DeepSeek-R1-Lite Preview is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after DeepSeek-R1-Lite Preview?** A: DeepSeek's previous tracked release was DeepSeek V2.5 on Sep 5 2024, 76 days earlier. It was followed by DeepSeek V2.5 (Revised) on Dec 10 2024. ### Mistral Large 2 24.11 - URL: https://aireleasetracker.com/model/mistral/mistral-large-2-24.11 - Developer: Mistral - Release date: Monday, Nov 18 2024 - Licensing: Open Weight - Benchmarks: none published in this dataset **Q: When was Mistral Large 2 24.11 released?** A: Mistral Large 2 24.11 was released by Mistral on Monday, Nov 18 2024. **Q: Who made Mistral Large 2 24.11?** A: Mistral Large 2 24.11 was built by Mistral. French AI company building open and commercial models. Founded in 2023 by former Meta and DeepMind researchers. **Q: Is Mistral Large 2 24.11 open source?** A: Partly. Mistral Large 2 24.11 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Mistral Large 2 24.11?** A: Mistral's previous tracked release was Ministral 3B on Oct 16 2024, 33 days earlier. It was followed by Pixtral Large 24.11 on Nov 18 2024. ### Pixtral Large 24.11 - URL: https://aireleasetracker.com/model/mistral/pixtral-large-24.11 - Developer: Mistral - Release date: Monday, Nov 18 2024 - Licensing: Open Weight - Benchmarks: none published in this dataset **Q: When was Pixtral Large 24.11 released?** A: Pixtral Large 24.11 was released by Mistral on Monday, Nov 18 2024. **Q: Who made Pixtral Large 24.11?** A: Pixtral Large 24.11 was built by Mistral. French AI company building open and commercial models. Founded in 2023 by former Meta and DeepMind researchers. **Q: Is Pixtral Large 24.11 open source?** A: Partly. Pixtral Large 24.11 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Pixtral Large 24.11?** A: Mistral's previous tracked release was Mistral Large 2 24.11 on Nov 18 2024. It was followed by Codestral 25.01 on Jan 13 2025. ### Qwen2.5-Coder - URL: https://aireleasetracker.com/model/qwen/qwen2.5-coder - Developer: Qwen - Release date: Tuesday, Nov 12 2024 - Licensing: Open Weight - Parameters: 32B - Context window: 128k - Benchmarks: none published in this dataset **Q: When was Qwen2.5-Coder released?** A: Qwen2.5-Coder was released by Qwen on Tuesday, Nov 12 2024. **Q: Who made Qwen2.5-Coder?** A: Qwen2.5-Coder was built by Qwen. Alibaba's AI lab, building the Qwen family. The most prolific publisher of open-weight models of any major lab, alongside a proprietary Max and Plus tier sold through Alibaba Cloud. **Q: What is the context window of Qwen2.5-Coder?** A: Qwen2.5-Coder has a context window of 128k. That is the maximum amount of input plus output the model can hold in a single request. **Q: How many parameters does Qwen2.5-Coder have?** A: Qwen2.5-Coder is reported at 32B parameters. **Q: Is Qwen2.5-Coder open source?** A: Partly. Qwen2.5-Coder is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Qwen2.5-Coder?** A: Qwen's previous tracked release was Qwen2.5 on Sep 19 2024, 54 days earlier. It was followed by QwQ-32B-Preview on Nov 28 2024. ### Claude 3.5 Haiku - URL: https://aireleasetracker.com/model/anthropic/claude-3.5-haiku - Developer: Anthropic - Release date: Tuesday, Oct 22 2024 - Licensing: Proprietary - Benchmarks published at release: - BullshitBench v2: 50% - SWE-Bench Verified: 40.6% - GPQA Diamond: 41.6% **Q: When was Claude 3.5 Haiku released?** A: Claude 3.5 Haiku was released by Anthropic on Tuesday, Oct 22 2024. **Q: Who made Claude 3.5 Haiku?** A: Claude 3.5 Haiku was built by Anthropic. AI safety company building the Claude family of models. Founded in 2021 by former OpenAI researchers. **Q: What benchmark scores did Claude 3.5 Haiku get?** A: Claude 3.5 Haiku reports 3 tracked benchmark scores — BullshitBench v2: 50%; SWE-Bench Verified: 40.6%; GPQA Diamond: 41.6%. Scores are the figures published at release by Anthropic. **Q: Is Claude 3.5 Haiku open source?** A: No. Claude 3.5 Haiku is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Claude 3.5 Haiku?** A: Anthropic's previous tracked release was Claude 3.5 Sonnet on Jun 20 2024, 124 days earlier. It was followed by Claude 3.5 Sonnet (upgraded) on Oct 22 2024. ### Claude 3.5 Sonnet (upgraded) - URL: https://aireleasetracker.com/model/anthropic/claude-3.5-sonnet-(upgraded) - Developer: Anthropic - Release date: Tuesday, Oct 22 2024 - Licensing: Proprietary - Benchmarks published at release: - SWE-Bench Verified: 49% **Q: When was Claude 3.5 Sonnet (upgraded) released?** A: Claude 3.5 Sonnet (upgraded) was released by Anthropic on Tuesday, Oct 22 2024. **Q: Who made Claude 3.5 Sonnet (upgraded)?** A: Claude 3.5 Sonnet (upgraded) was built by Anthropic. AI safety company building the Claude family of models. Founded in 2021 by former OpenAI researchers. **Q: What benchmark scores did Claude 3.5 Sonnet (upgraded) get?** A: Claude 3.5 Sonnet (upgraded) reports 1 tracked benchmark score — SWE-Bench Verified: 49%. Scores are the figures published at release by Anthropic. **Q: Is Claude 3.5 Sonnet (upgraded) open source?** A: No. Claude 3.5 Sonnet (upgraded) is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Claude 3.5 Sonnet (upgraded)?** A: Anthropic's previous tracked release was Claude 3.5 Haiku on Oct 22 2024. It was followed by Claude 3.7 Sonnet on Feb 24 2025. ### Ministral 8B - URL: https://aireleasetracker.com/model/mistral/ministral-8b - Developer: Mistral - Release date: Wednesday, Oct 16 2024 - Licensing: Open Weight - Benchmarks: none published in this dataset **Q: When was Ministral 8B released?** A: Ministral 8B was released by Mistral on Wednesday, Oct 16 2024. **Q: Who made Ministral 8B?** A: Ministral 8B was built by Mistral. French AI company building open and commercial models. Founded in 2023 by former Meta and DeepMind researchers. **Q: Is Ministral 8B open source?** A: Partly. Ministral 8B is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Ministral 8B?** A: Mistral's previous tracked release was Pixtral 24.09 on Sep 11 2024, 35 days earlier. It was followed by Ministral 3B on Oct 16 2024. ### Ministral 3B - URL: https://aireleasetracker.com/model/mistral/ministral-3b - Developer: Mistral - Release date: Wednesday, Oct 16 2024 - Licensing: Open Weight - Benchmarks: none published in this dataset **Q: When was Ministral 3B released?** A: Ministral 3B was released by Mistral on Wednesday, Oct 16 2024. **Q: Who made Ministral 3B?** A: Ministral 3B was built by Mistral. French AI company building open and commercial models. Founded in 2023 by former Meta and DeepMind researchers. **Q: Is Ministral 3B open source?** A: Partly. Ministral 3B is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Ministral 3B?** A: Mistral's previous tracked release was Ministral 8B on Oct 16 2024. It was followed by Mistral Large 2 24.11 on Nov 18 2024. ### Gemini 1.5 Flash-8B - URL: https://aireleasetracker.com/model/google/gemini-1.5-flash-8b - Developer: Google - Release date: Thursday, Oct 3 2024 - Licensing: Proprietary - Benchmarks: none published in this dataset **Q: When was Gemini 1.5 Flash-8B released?** A: Gemini 1.5 Flash-8B was released by Google on Thursday, Oct 3 2024. **Q: Who made Gemini 1.5 Flash-8B?** A: Gemini 1.5 Flash-8B was built by Google. Builds the Gemini family of models through Google DeepMind. Integrates AI across Google products. **Q: Is Gemini 1.5 Flash-8B open source?** A: No. Gemini 1.5 Flash-8B is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Gemini 1.5 Flash-8B?** A: Google's previous tracked release was Gemini 1.5 Flash-002 on Sep 24 2024, 9 days earlier. It was followed by Gemini 2.0 Flash on Jan 30 2025. ### LLaMA 3.2 - URL: https://aireleasetracker.com/model/meta/llama-3.2 - Developer: Meta - Release date: Wednesday, Sep 25 2024 - Licensing: Open Weight - Parameters: 1B/3B - Benchmarks: none published in this dataset **Q: When was LLaMA 3.2 released?** A: LLaMA 3.2 was released by Meta on Wednesday, Sep 25 2024. **Q: Who made LLaMA 3.2?** A: LLaMA 3.2 was built by Meta. Develops the open-weight Llama series of models. Committed to open-source AI research. **Q: How many parameters does LLaMA 3.2 have?** A: LLaMA 3.2 is reported at 1B/3B parameters. **Q: Is LLaMA 3.2 open source?** A: Partly. LLaMA 3.2 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after LLaMA 3.2?** A: Meta's previous tracked release was LLaMA 3.1 on Jul 23 2024, 64 days earlier. It was followed by LLaMA 3.3 on Dec 6 2024. ### Gemini 1.5 Pro-002 - URL: https://aireleasetracker.com/model/google/gemini-1.5-pro-002 - Developer: Google - Release date: Tuesday, Sep 24 2024 - Licensing: Proprietary - Benchmarks: none published in this dataset **Q: When was Gemini 1.5 Pro-002 released?** A: Gemini 1.5 Pro-002 was released by Google on Tuesday, Sep 24 2024. **Q: Who made Gemini 1.5 Pro-002?** A: Gemini 1.5 Pro-002 was built by Google. Builds the Gemini family of models through Google DeepMind. Integrates AI across Google products. **Q: Is Gemini 1.5 Pro-002 open source?** A: No. Gemini 1.5 Pro-002 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Gemini 1.5 Pro-002?** A: Google's previous tracked release was Gemma 2 on Jun 27 2024, 89 days earlier. It was followed by Gemini 1.5 Flash-002 on Sep 24 2024. ### Gemini 1.5 Flash-002 - URL: https://aireleasetracker.com/model/google/gemini-1.5-flash-002 - Developer: Google - Release date: Tuesday, Sep 24 2024 - Licensing: Proprietary - Benchmarks: none published in this dataset **Q: When was Gemini 1.5 Flash-002 released?** A: Gemini 1.5 Flash-002 was released by Google on Tuesday, Sep 24 2024. **Q: Who made Gemini 1.5 Flash-002?** A: Gemini 1.5 Flash-002 was built by Google. Builds the Gemini family of models through Google DeepMind. Integrates AI across Google products. **Q: Is Gemini 1.5 Flash-002 open source?** A: No. Gemini 1.5 Flash-002 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Gemini 1.5 Flash-002?** A: Google's previous tracked release was Gemini 1.5 Pro-002 on Sep 24 2024. It was followed by Gemini 1.5 Flash-8B on Oct 3 2024. ### Qwen2.5 - URL: https://aireleasetracker.com/model/qwen/qwen2.5 - Developer: Qwen - Release date: Thursday, Sep 19 2024 - Licensing: Open Weight - Parameters: 72B - Context window: 128k - Benchmarks: none published in this dataset Qwen2.5, released September 19, 2024, was the generation that made Qwen the default open-weight base model for the wider research community. It filled in the size ladder with 3B, 14B and 32B alongside the existing 0.5B, 1.5B, 7B and 72B, so there was a Qwen checkpoint sized for almost any deployment budget, and it shipped with a 128K-token context window across the line. Licensing was the other half of the story: everything except the 72B flagship and the 3B model went out under Apache 2.0, which made Qwen2.5 far easier to build commercial products on than the Llama licence allowed. Alibaba released Qwen2.5-Coder and Qwen2.5-Math the same day, and by the time Qwen3 arrived the following April, Qwen2.5 derivatives and fine-tunes outnumbered those of any other base model family on Hugging Face. **Q: When was Qwen2.5 released?** A: Qwen2.5 was released by Qwen on Thursday, Sep 19 2024. **Q: Who made Qwen2.5?** A: Qwen2.5 was built by Qwen. Alibaba's AI lab, building the Qwen family. The most prolific publisher of open-weight models of any major lab, alongside a proprietary Max and Plus tier sold through Alibaba Cloud. **Q: What is the context window of Qwen2.5?** A: Qwen2.5 has a context window of 128k. That is the maximum amount of input plus output the model can hold in a single request. **Q: How many parameters does Qwen2.5 have?** A: Qwen2.5 is reported at 72B parameters. **Q: Is Qwen2.5 open source?** A: Partly. Qwen2.5 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Qwen2.5?** A: Qwen's previous tracked release was Qwen2 on Jun 6 2024, 105 days earlier. It was followed by Qwen2.5-Coder on Nov 12 2024. ### o1-preview - URL: https://aireleasetracker.com/model/openai/o1-preview - Developer: OpenAI - Release date: Thursday, Sep 12 2024 - Licensing: Proprietary - Context window: 128k - Benchmarks published at release: - GPQA Diamond: 73.3% **Q: When was o1-preview released?** A: o1-preview was released by OpenAI on Thursday, Sep 12 2024. **Q: Who made o1-preview?** A: o1-preview was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: What benchmark scores did o1-preview get?** A: o1-preview reports 1 tracked benchmark score — GPQA Diamond: 73.3%. Scores are the figures published at release by OpenAI. **Q: What is the context window of o1-preview?** A: o1-preview has a context window of 128k. That is the maximum amount of input plus output the model can hold in a single request. **Q: Is o1-preview open source?** A: No. o1-preview is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after o1-preview?** A: OpenAI's previous tracked release was GPT-4o mini on Jul 18 2024, 56 days earlier. It was followed by o1-mini on Sep 12 2024. ### o1-mini - URL: https://aireleasetracker.com/model/openai/o1-mini - Developer: OpenAI - Release date: Thursday, Sep 12 2024 - Licensing: Proprietary - Context window: 128k - Benchmarks published at release: - GPQA Diamond: 60% **Q: When was o1-mini released?** A: o1-mini was released by OpenAI on Thursday, Sep 12 2024. **Q: Who made o1-mini?** A: o1-mini was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: What benchmark scores did o1-mini get?** A: o1-mini reports 1 tracked benchmark score — GPQA Diamond: 60%. Scores are the figures published at release by OpenAI. **Q: What is the context window of o1-mini?** A: o1-mini has a context window of 128k. That is the maximum amount of input plus output the model can hold in a single request. **Q: Is o1-mini open source?** A: No. o1-mini is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after o1-mini?** A: OpenAI's previous tracked release was o1-preview on Sep 12 2024. It was followed by o1 on Dec 5 2024. ### Pixtral 24.09 - URL: https://aireleasetracker.com/model/mistral/pixtral-24.09 - Developer: Mistral - Release date: Wednesday, Sep 11 2024 - Licensing: Open Weight - Benchmarks: none published in this dataset **Q: When was Pixtral 24.09 released?** A: Pixtral 24.09 was released by Mistral on Wednesday, Sep 11 2024. **Q: Who made Pixtral 24.09?** A: Pixtral 24.09 was built by Mistral. French AI company building open and commercial models. Founded in 2023 by former Meta and DeepMind researchers. **Q: Is Pixtral 24.09 open source?** A: Partly. Pixtral 24.09 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Pixtral 24.09?** A: Mistral's previous tracked release was Mistral Large 2 on Jul 24 2024, 49 days earlier. It was followed by Ministral 8B on Oct 16 2024. ### DeepSeek V2.5 - URL: https://aireleasetracker.com/model/deepseek/deepseek-v2.5 - Developer: DeepSeek - Release date: Thursday, Sep 5 2024 - Licensing: Open Weight - Benchmarks: none published in this dataset **Q: When was DeepSeek V2.5 released?** A: DeepSeek V2.5 was released by DeepSeek on Thursday, Sep 5 2024. **Q: Who made DeepSeek V2.5?** A: DeepSeek V2.5 was built by DeepSeek. Chinese AI lab known for efficient, open-weight models. Gained attention for strong performance at lower cost. **Q: Is DeepSeek V2.5 open source?** A: Partly. DeepSeek V2.5 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after DeepSeek V2.5?** A: DeepSeek's previous tracked release was DeepSeek Coder V2 on Jun 17 2024, 80 days earlier. It was followed by DeepSeek-R1-Lite Preview on Nov 20 2024. ### GLM-4-Plus - URL: https://aireleasetracker.com/model/zai/glm-4-plus - Developer: Z.ai - Release date: Thursday, Aug 29 2024 - Licensing: Proprietary - Benchmarks: none published in this dataset **Q: When was GLM-4-Plus released?** A: GLM-4-Plus was released by Z.ai on Thursday, Aug 29 2024. **Q: Who made GLM-4-Plus?** A: GLM-4-Plus was built by Z.ai. Chinese AI lab spun out of Tsinghua University (formerly Zhipu AI), building the open-weight GLM family. Rebranded internationally as Z.ai in 2025. **Q: Is GLM-4-Plus open source?** A: No. GLM-4-Plus is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after GLM-4-Plus?** A: Z.ai's previous tracked release was GLM-4-9B on Jun 5 2024, 85 days earlier. It was followed by GLM-4.5 on Jul 28 2025. ### Grok‑2 - URL: https://aireleasetracker.com/model/xai/grok‑2 - Developer: SpaceXAI - Release date: Wednesday, Aug 14 2024 - Licensing: Proprietary - Benchmarks: none published in this dataset **Q: When was Grok‑2 released?** A: Grok‑2 was released by SpaceXAI on Wednesday, Aug 14 2024. **Q: Who made Grok‑2?** A: Grok‑2 was built by SpaceXAI. Elon Musk's AI company building the Grok series of models. Founded in 2023 as xAI, now part of SpaceX. **Q: Is Grok‑2 open source?** A: No. Grok‑2 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Grok‑2?** A: SpaceXAI's previous tracked release was Grok‑1.5 on Mar 29 2024, 138 days earlier. It was followed by Grok 3 on Feb 17 2025. ### Mistral Large 2 - URL: https://aireleasetracker.com/model/mistral/mistral-large-2 - Developer: Mistral - Release date: Wednesday, Jul 24 2024 - Licensing: Open Weight - Benchmarks: none published in this dataset **Q: When was Mistral Large 2 released?** A: Mistral Large 2 was released by Mistral on Wednesday, Jul 24 2024. **Q: Who made Mistral Large 2?** A: Mistral Large 2 was built by Mistral. French AI company building open and commercial models. Founded in 2023 by former Meta and DeepMind researchers. **Q: Is Mistral Large 2 open source?** A: Partly. Mistral Large 2 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Mistral Large 2?** A: Mistral's previous tracked release was Mistral NeMo on Jul 18 2024, 6 days earlier. It was followed by Pixtral 24.09 on Sep 11 2024. ### LLaMA 3.1 - URL: https://aireleasetracker.com/model/meta/llama-3.1 - Developer: Meta - Release date: Tuesday, Jul 23 2024 - Licensing: Open Weight - Benchmarks published at release: - BullshitBench v2: 14% Llama 3.1, released July 23, 2024, delivered the first open-weight frontier-class model: the 405B variant traded blows with GPT-4o and Claude 3.5 Sonnet on standard benchmarks while being free to download. The 8B and 70B sizes were refreshed alongside it, and the whole family gained a 128K context window and stronger multilingual coverage. Meta paired the release with an updated license permitting synthetic-data generation — making 405B a popular teacher model for distillation — and an open letter from Mark Zuckerberg arguing open-source AI was the path forward. It stood as the open-weights capability ceiling until DeepSeek V3 arrived that December. **Q: When was LLaMA 3.1 released?** A: LLaMA 3.1 was released by Meta on Tuesday, Jul 23 2024. **Q: Who made LLaMA 3.1?** A: LLaMA 3.1 was built by Meta. Develops the open-weight Llama series of models. Committed to open-source AI research. **Q: What benchmark scores did LLaMA 3.1 get?** A: LLaMA 3.1 reports 1 tracked benchmark score — BullshitBench v2: 14%. Scores are the figures published at release by Meta. **Q: Is LLaMA 3.1 open source?** A: Partly. LLaMA 3.1 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after LLaMA 3.1?** A: Meta's previous tracked release was LLaMA 3 (8B/70B) on Apr 18 2024, 96 days earlier. It was followed by LLaMA 3.2 on Sep 25 2024. ### GPT-4o mini - URL: https://aireleasetracker.com/model/openai/gpt-4o-mini - Developer: OpenAI - Release date: Thursday, Jul 18 2024 - Licensing: Proprietary - Context window: 128k - Benchmarks published at release: - BullshitBench v2: 2% - GPQA Diamond: 40.2% **Q: When was GPT-4o mini released?** A: GPT-4o mini was released by OpenAI on Thursday, Jul 18 2024. **Q: Who made GPT-4o mini?** A: GPT-4o mini was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: What benchmark scores did GPT-4o mini get?** A: GPT-4o mini reports 2 tracked benchmark scores — BullshitBench v2: 2%; GPQA Diamond: 40.2%. Scores are the figures published at release by OpenAI. **Q: What is the context window of GPT-4o mini?** A: GPT-4o mini has a context window of 128k. That is the maximum amount of input plus output the model can hold in a single request. **Q: Is GPT-4o mini open source?** A: No. GPT-4o mini is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after GPT-4o mini?** A: OpenAI's previous tracked release was GPT-4o on May 13 2024, 66 days earlier. It was followed by o1-preview on Sep 12 2024. ### Mistral NeMo - URL: https://aireleasetracker.com/model/mistral/mistral-nemo - Developer: Mistral - Release date: Thursday, Jul 18 2024 - Licensing: Open Weight - Parameters: 12B - Context window: 128k - Benchmarks: none published in this dataset **Q: When was Mistral NeMo released?** A: Mistral NeMo was released by Mistral on Thursday, Jul 18 2024. **Q: Who made Mistral NeMo?** A: Mistral NeMo was built by Mistral. French AI company building open and commercial models. Founded in 2023 by former Meta and DeepMind researchers. **Q: What is the context window of Mistral NeMo?** A: Mistral NeMo has a context window of 128k. That is the maximum amount of input plus output the model can hold in a single request. **Q: How many parameters does Mistral NeMo have?** A: Mistral NeMo is reported at 12B parameters. **Q: Is Mistral NeMo open source?** A: Partly. Mistral NeMo is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Mistral NeMo?** A: Mistral's previous tracked release was Mathstral 7B on Jul 16 2024, 2 days earlier. It was followed by Mistral Large 2 on Jul 24 2024. ### Codestral Mamba 7B - URL: https://aireleasetracker.com/model/mistral/codestral-mamba-7b - Developer: Mistral - Release date: Tuesday, Jul 16 2024 - Licensing: Open Weight - Benchmarks: none published in this dataset **Q: When was Codestral Mamba 7B released?** A: Codestral Mamba 7B was released by Mistral on Tuesday, Jul 16 2024. **Q: Who made Codestral Mamba 7B?** A: Codestral Mamba 7B was built by Mistral. French AI company building open and commercial models. Founded in 2023 by former Meta and DeepMind researchers. **Q: Is Codestral Mamba 7B open source?** A: Partly. Codestral Mamba 7B is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Codestral Mamba 7B?** A: Mistral's previous tracked release was Codestral 22B on May 29 2024, 48 days earlier. It was followed by Mathstral 7B on Jul 16 2024. ### Mathstral 7B - URL: https://aireleasetracker.com/model/mistral/mathstral-7b - Developer: Mistral - Release date: Tuesday, Jul 16 2024 - Licensing: Open Weight - Benchmarks: none published in this dataset **Q: When was Mathstral 7B released?** A: Mathstral 7B was released by Mistral on Tuesday, Jul 16 2024. **Q: Who made Mathstral 7B?** A: Mathstral 7B was built by Mistral. French AI company building open and commercial models. Founded in 2023 by former Meta and DeepMind researchers. **Q: Is Mathstral 7B open source?** A: Partly. Mathstral 7B is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Mathstral 7B?** A: Mistral's previous tracked release was Codestral Mamba 7B on Jul 16 2024. It was followed by Mistral NeMo on Jul 18 2024. ### Gemma 2 - URL: https://aireleasetracker.com/model/google/gemma-2 - Developer: Google - Release date: Thursday, Jun 27 2024 - Licensing: Open Weight - Benchmarks: none published in this dataset **Q: When was Gemma 2 released?** A: Gemma 2 was released by Google on Thursday, Jun 27 2024. **Q: Who made Gemma 2?** A: Gemma 2 was built by Google. Builds the Gemini family of models through Google DeepMind. Integrates AI across Google products. **Q: Is Gemma 2 open source?** A: Partly. Gemma 2 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Gemma 2?** A: Google's previous tracked release was Gemini 1.5 Flash on May 14 2024, 44 days earlier. It was followed by Gemini 1.5 Pro-002 on Sep 24 2024. ### Claude 3.5 Sonnet - URL: https://aireleasetracker.com/model/anthropic/claude-3.5-sonnet - Developer: Anthropic - Release date: Thursday, Jun 20 2024 - Licensing: Proprietary - Benchmarks published at release: - BullshitBench v2: 45% - SWE-Bench Verified: 33.4% - GPQA Diamond: 59.4% Claude 3.5 Sonnet, released June 20, 2024, was the model that made Claude a default choice for developers. Despite being a mid-tier model, it outperformed the larger Claude 3 Opus while running roughly twice as fast at a fraction of the cost, and its coding and vision abilities set a new bar for the price point. It launched together with Artifacts in the Claude app, which turned chat output into live, editable documents and code. On benchmarks it jumped to 59.4% on GPQA Diamond — nine points over Opus — and it became the base of one of the most-used model lines in AI-assisted programming. An upgraded version followed on October 22, 2024, lifting SWE-Bench Verified to 49.0% and introducing computer use, before Claude 3.7 Sonnet succeeded the line in February 2025. **Q: When was Claude 3.5 Sonnet released?** A: Claude 3.5 Sonnet was released by Anthropic on Thursday, Jun 20 2024. **Q: Who made Claude 3.5 Sonnet?** A: Claude 3.5 Sonnet was built by Anthropic. AI safety company building the Claude family of models. Founded in 2021 by former OpenAI researchers. **Q: What benchmark scores did Claude 3.5 Sonnet get?** A: Claude 3.5 Sonnet reports 3 tracked benchmark scores — BullshitBench v2: 45%; SWE-Bench Verified: 33.4%; GPQA Diamond: 59.4%. Scores are the figures published at release by Anthropic. **Q: Is Claude 3.5 Sonnet open source?** A: No. Claude 3.5 Sonnet is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Claude 3.5 Sonnet?** A: Anthropic's previous tracked release was Claude 3 Opus on Mar 4 2024, 108 days earlier. It was followed by Claude 3.5 Haiku on Oct 22 2024. ### DeepSeek Coder V2 - URL: https://aireleasetracker.com/model/deepseek/deepseek-coder-v2 - Developer: DeepSeek - Release date: Monday, Jun 17 2024 - Licensing: Open Weight - Benchmarks: none published in this dataset **Q: When was DeepSeek Coder V2 released?** A: DeepSeek Coder V2 was released by DeepSeek on Monday, Jun 17 2024. **Q: Who made DeepSeek Coder V2?** A: DeepSeek Coder V2 was built by DeepSeek. Chinese AI lab known for efficient, open-weight models. Gained attention for strong performance at lower cost. **Q: Is DeepSeek Coder V2 open source?** A: Partly. DeepSeek Coder V2 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after DeepSeek Coder V2?** A: DeepSeek's previous tracked release was DeepSeek V2 on May 6 2024, 42 days earlier. It was followed by DeepSeek V2.5 on Sep 5 2024. ### Qwen2 - URL: https://aireleasetracker.com/model/qwen/qwen2 - Developer: Qwen - Release date: Thursday, Jun 6 2024 - Licensing: Open Weight - Parameters: 72B - Context window: 128k - Benchmarks: none published in this dataset **Q: When was Qwen2 released?** A: Qwen2 was released by Qwen on Thursday, Jun 6 2024. **Q: Who made Qwen2?** A: Qwen2 was built by Qwen. Alibaba's AI lab, building the Qwen family. The most prolific publisher of open-weight models of any major lab, alongside a proprietary Max and Plus tier sold through Alibaba Cloud. **Q: What is the context window of Qwen2?** A: Qwen2 has a context window of 128k. That is the maximum amount of input plus output the model can hold in a single request. **Q: How many parameters does Qwen2 have?** A: Qwen2 is reported at 72B parameters. **Q: Is Qwen2 open source?** A: Partly. Qwen2 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Qwen2?** A: Qwen's previous tracked release was Qwen1.5 on Feb 4 2024, 123 days earlier. It was followed by Qwen2.5 on Sep 19 2024. ### GLM-4-9B - URL: https://aireleasetracker.com/model/zai/glm-4-9b - Developer: Z.ai - Release date: Wednesday, Jun 5 2024 - Licensing: Open Weight - Parameters: 9B - Benchmarks: none published in this dataset **Q: When was GLM-4-9B released?** A: GLM-4-9B was released by Z.ai on Wednesday, Jun 5 2024. **Q: Who made GLM-4-9B?** A: GLM-4-9B was built by Z.ai. Chinese AI lab spun out of Tsinghua University (formerly Zhipu AI), building the open-weight GLM family. Rebranded internationally as Z.ai in 2025. **Q: How many parameters does GLM-4-9B have?** A: GLM-4-9B is reported at 9B parameters. **Q: Is GLM-4-9B open source?** A: Partly. GLM-4-9B is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after GLM-4-9B?** A: Z.ai's previous tracked release was GLM-4 on Jan 16 2024, 141 days earlier. It was followed by GLM-4-Plus on Aug 29 2024. ### Codestral 22B - URL: https://aireleasetracker.com/model/mistral/codestral-22b - Developer: Mistral - Release date: Wednesday, May 29 2024 - Licensing: Proprietary - Benchmarks: none published in this dataset **Q: When was Codestral 22B released?** A: Codestral 22B was released by Mistral on Wednesday, May 29 2024. **Q: Who made Codestral 22B?** A: Codestral 22B was built by Mistral. French AI company building open and commercial models. Founded in 2023 by former Meta and DeepMind researchers. **Q: Is Codestral 22B open source?** A: No. Codestral 22B is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Codestral 22B?** A: Mistral's previous tracked release was Mixtral 8×22B on Apr 10 2024, 49 days earlier. It was followed by Codestral Mamba 7B on Jul 16 2024. ### Gemini 1.5 Flash - URL: https://aireleasetracker.com/model/google/gemini-1.5-flash - Developer: Google - Release date: Tuesday, May 14 2024 - Licensing: Proprietary - Benchmarks: none published in this dataset **Q: When was Gemini 1.5 Flash released?** A: Gemini 1.5 Flash was released by Google on Tuesday, May 14 2024. **Q: Who made Gemini 1.5 Flash?** A: Gemini 1.5 Flash was built by Google. Builds the Gemini family of models through Google DeepMind. Integrates AI across Google products. **Q: Is Gemini 1.5 Flash open source?** A: No. Gemini 1.5 Flash is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Gemini 1.5 Flash?** A: Google's previous tracked release was Gemma 1 on Feb 21 2024, 83 days earlier. It was followed by Gemma 2 on Jun 27 2024. ### GPT-4o - URL: https://aireleasetracker.com/model/openai/gpt-4o - Developer: OpenAI - Release date: Monday, May 13 2024 - Licensing: Proprietary - Context window: 128k - Benchmarks published at release: - BullshitBench v2: 12% - GPQA Diamond: 49.9% GPT-4o — the "o" stands for omni — launched May 13, 2024 as OpenAI's first natively multimodal model, trained end-to-end across text, vision, and audio rather than stitching separate models together. Its headline feature was real-time voice conversation with near-human response latency, demoed live the day before Google I/O. GPT-4o also marked a strategic shift: it brought GPT-4-class intelligence to ChatGPT's free tier for the first time, at half the API price of GPT-4 Turbo. The smaller GPT-4o mini followed in July 2024 and displaced GPT-3.5 as the default cheap model. GPT-4o remained ChatGPT's workhorse until the GPT-5 launch in August 2025. **Q: When was GPT-4o released?** A: GPT-4o was released by OpenAI on Monday, May 13 2024. **Q: Who made GPT-4o?** A: GPT-4o was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: What benchmark scores did GPT-4o get?** A: GPT-4o reports 2 tracked benchmark scores — BullshitBench v2: 12%; GPQA Diamond: 49.9%. Scores are the figures published at release by OpenAI. **Q: What is the context window of GPT-4o?** A: GPT-4o has a context window of 128k. That is the maximum amount of input plus output the model can hold in a single request. **Q: Is GPT-4o open source?** A: No. GPT-4o is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after GPT-4o?** A: OpenAI's previous tracked release was GPT-4 Turbo on Nov 6 2023, 189 days earlier. It was followed by GPT-4o mini on Jul 18 2024. ### DeepSeek V2 - URL: https://aireleasetracker.com/model/deepseek/deepseek-v2 - Developer: DeepSeek - Release date: Monday, May 6 2024 - Licensing: Open Weight - Benchmarks: none published in this dataset **Q: When was DeepSeek V2 released?** A: DeepSeek V2 was released by DeepSeek on Monday, May 6 2024. **Q: Who made DeepSeek V2?** A: DeepSeek V2 was built by DeepSeek. Chinese AI lab known for efficient, open-weight models. Gained attention for strong performance at lower cost. **Q: Is DeepSeek V2 open source?** A: Partly. DeepSeek V2 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after DeepSeek V2?** A: DeepSeek's previous tracked release was DeepSeek-Math on Feb 5 2024, 91 days earlier. It was followed by DeepSeek Coder V2 on Jun 17 2024. ### LLaMA 3 (8B/70B) - URL: https://aireleasetracker.com/model/meta/llama-3-(8b/70b) - Developer: Meta - Release date: Thursday, Apr 18 2024 - Licensing: Open Weight - Benchmarks: none published in this dataset **Q: When was LLaMA 3 (8B/70B) released?** A: LLaMA 3 (8B/70B) was released by Meta on Thursday, Apr 18 2024. **Q: Who made LLaMA 3 (8B/70B)?** A: LLaMA 3 (8B/70B) was built by Meta. Develops the open-weight Llama series of models. Committed to open-source AI research. **Q: Is LLaMA 3 (8B/70B) open source?** A: Partly. LLaMA 3 (8B/70B) is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after LLaMA 3 (8B/70B)?** A: Meta's previous tracked release was Code Llama 70B on Jan 29 2024, 80 days earlier. It was followed by LLaMA 3.1 on Jul 23 2024. ### Mixtral 8×22B - URL: https://aireleasetracker.com/model/mistral/mixtral-8×22b - Developer: Mistral - Release date: Wednesday, Apr 10 2024 - Licensing: Open Weight - Benchmarks: none published in this dataset **Q: When was Mixtral 8×22B released?** A: Mixtral 8×22B was released by Mistral on Wednesday, Apr 10 2024. **Q: Who made Mixtral 8×22B?** A: Mixtral 8×22B was built by Mistral. French AI company building open and commercial models. Founded in 2023 by former Meta and DeepMind researchers. **Q: Is Mixtral 8×22B open source?** A: Partly. Mixtral 8×22B is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Mixtral 8×22B?** A: Mistral's previous tracked release was Mistral Small on Feb 26 2024, 44 days earlier. It was followed by Codestral 22B on May 29 2024. ### Grok‑1.5 - URL: https://aireleasetracker.com/model/xai/grok‑1.5 - Developer: SpaceXAI - Release date: Friday, Mar 29 2024 - Licensing: Proprietary - Benchmarks: none published in this dataset **Q: When was Grok‑1.5 released?** A: Grok‑1.5 was released by SpaceXAI on Friday, Mar 29 2024. **Q: Who made Grok‑1.5?** A: Grok‑1.5 was built by SpaceXAI. Elon Musk's AI company building the Grok series of models. Founded in 2023 as xAI, now part of SpaceX. **Q: Is Grok‑1.5 open source?** A: No. Grok‑1.5 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Grok‑1.5?** A: SpaceXAI's previous tracked release was Grok‑1 on Nov 3 2023, 147 days earlier. It was followed by Grok‑2 on Aug 14 2024. ### Claude 3 Haiku - URL: https://aireleasetracker.com/model/anthropic/claude-3-haiku - Developer: Anthropic - Release date: Monday, Mar 4 2024 - Licensing: Proprietary - Benchmarks published at release: - BullshitBench v2: 10% - GPQA Diamond: 33.3% **Q: When was Claude 3 Haiku released?** A: Claude 3 Haiku was released by Anthropic on Monday, Mar 4 2024. **Q: Who made Claude 3 Haiku?** A: Claude 3 Haiku was built by Anthropic. AI safety company building the Claude family of models. Founded in 2021 by former OpenAI researchers. **Q: What benchmark scores did Claude 3 Haiku get?** A: Claude 3 Haiku reports 2 tracked benchmark scores — BullshitBench v2: 10%; GPQA Diamond: 33.3%. Scores are the figures published at release by Anthropic. **Q: Is Claude 3 Haiku open source?** A: No. Claude 3 Haiku is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Claude 3 Haiku?** A: Anthropic's previous tracked release was Claude 2.1 on Nov 21 2023, 104 days earlier. It was followed by Claude 3 Sonnet on Mar 4 2024. ### Claude 3 Sonnet - URL: https://aireleasetracker.com/model/anthropic/claude-3-sonnet - Developer: Anthropic - Release date: Monday, Mar 4 2024 - Licensing: Proprietary - Benchmarks published at release: - GPQA Diamond: 40.4% **Q: When was Claude 3 Sonnet released?** A: Claude 3 Sonnet was released by Anthropic on Monday, Mar 4 2024. **Q: Who made Claude 3 Sonnet?** A: Claude 3 Sonnet was built by Anthropic. AI safety company building the Claude family of models. Founded in 2021 by former OpenAI researchers. **Q: What benchmark scores did Claude 3 Sonnet get?** A: Claude 3 Sonnet reports 1 tracked benchmark score — GPQA Diamond: 40.4%. Scores are the figures published at release by Anthropic. **Q: Is Claude 3 Sonnet open source?** A: No. Claude 3 Sonnet is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Claude 3 Sonnet?** A: Anthropic's previous tracked release was Claude 3 Haiku on Mar 4 2024. It was followed by Claude 3 Opus on Mar 4 2024. ### Claude 3 Opus - URL: https://aireleasetracker.com/model/anthropic/claude-3-opus - Developer: Anthropic - Release date: Monday, Mar 4 2024 - Licensing: Proprietary - Benchmarks published at release: - SWE-Bench Verified: 33% - GPQA Diamond: 50.4% Claude 3 Opus was the flagship of the Claude 3 family, launched on March 4, 2024 alongside the faster Claude 3 Sonnet and Claude 3 Haiku. It was the first Claude generation with vision support, shipped with a 200K-token context window, and outperformed GPT-4 on a wide range of benchmarks at launch — a moment that established Anthropic as a genuine frontier-lab rival rather than a safety-focused also-ran. Opus was marketed on graduate-level reasoning: it scored 50.4% on GPQA Diamond and 33.0% on SWE-Bench Verified, both leading numbers for early 2024. It remained Anthropic's largest model for over a year, until the Claude 4 generation arrived in May 2025, and its three-tier Haiku/Sonnet/Opus naming scheme still structures the Claude lineup today. **Q: When was Claude 3 Opus released?** A: Claude 3 Opus was released by Anthropic on Monday, Mar 4 2024. **Q: Who made Claude 3 Opus?** A: Claude 3 Opus was built by Anthropic. AI safety company building the Claude family of models. Founded in 2021 by former OpenAI researchers. **Q: What benchmark scores did Claude 3 Opus get?** A: Claude 3 Opus reports 2 tracked benchmark scores — SWE-Bench Verified: 33%; GPQA Diamond: 50.4%. Scores are the figures published at release by Anthropic. **Q: Is Claude 3 Opus open source?** A: No. Claude 3 Opus is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Claude 3 Opus?** A: Anthropic's previous tracked release was Claude 3 Sonnet on Mar 4 2024. It was followed by Claude 3.5 Sonnet on Jun 20 2024. ### Mistral Large - URL: https://aireleasetracker.com/model/mistral/mistral-large - Developer: Mistral - Release date: Monday, Feb 26 2024 - Licensing: Proprietary - Benchmarks: none published in this dataset Mistral Large, released February 26, 2024 alongside Mistral Small, was the company's pivot to a commercial frontier business: a closed flagship API model, second only to GPT-4 on several benchmarks at launch, with strong native fluency across French, German, Spanish, and Italian and a distribution partnership with Microsoft Azure. The dual-track strategy it inaugurated — open-weight releases like Mixtral alongside proprietary flagships — became Mistral's signature. Mistral Large 2 (July 2024) returned the flagship tier to open weights at 123B parameters, and the line continued through Mistral Large 3 in December 2025. **Q: When was Mistral Large released?** A: Mistral Large was released by Mistral on Monday, Feb 26 2024. **Q: Who made Mistral Large?** A: Mistral Large was built by Mistral. French AI company building open and commercial models. Founded in 2023 by former Meta and DeepMind researchers. **Q: Is Mistral Large open source?** A: No. Mistral Large is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Mistral Large?** A: Mistral's previous tracked release was Mistral Medium on Dec 11 2023, 77 days earlier. It was followed by Mistral Small on Feb 26 2024. ### Mistral Small - URL: https://aireleasetracker.com/model/mistral/mistral-small - Developer: Mistral - Release date: Monday, Feb 26 2024 - Licensing: Proprietary - Benchmarks: none published in this dataset **Q: When was Mistral Small released?** A: Mistral Small was released by Mistral on Monday, Feb 26 2024. **Q: Who made Mistral Small?** A: Mistral Small was built by Mistral. French AI company building open and commercial models. Founded in 2023 by former Meta and DeepMind researchers. **Q: Is Mistral Small open source?** A: No. Mistral Small is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Mistral Small?** A: Mistral's previous tracked release was Mistral Large on Feb 26 2024. It was followed by Mixtral 8×22B on Apr 10 2024. ### Gemma 1 - URL: https://aireleasetracker.com/model/google/gemma-1 - Developer: Google - Release date: Wednesday, Feb 21 2024 - Licensing: Open Weight - Benchmarks: none published in this dataset **Q: When was Gemma 1 released?** A: Gemma 1 was released by Google on Wednesday, Feb 21 2024. **Q: Who made Gemma 1?** A: Gemma 1 was built by Google. Builds the Gemini family of models through Google DeepMind. Integrates AI across Google products. **Q: Is Gemma 1 open source?** A: Partly. Gemma 1 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Gemma 1?** A: Google's previous tracked release was Gemini 1.5 Pro on Feb 15 2024, 6 days earlier. It was followed by Gemini 1.5 Flash on May 14 2024. ### Gemini 1.5 Pro - URL: https://aireleasetracker.com/model/google/gemini-1.5-pro - Developer: Google - Release date: Thursday, Feb 15 2024 - Licensing: Proprietary - Benchmarks: none published in this dataset Gemini 1.5 Pro, announced February 15, 2024 — just a week after Gemini 1.0 Ultra — made long context the story of 2024. Built on a mixture-of-experts architecture, it shipped with a 1M-token context window (later 2M), an order of magnitude beyond anything else available, and could ingest hour-long videos, entire codebases, or novels in a single prompt. Its needle-in-a-haystack recall results forced every competitor to race toward longer contexts, and its Flash sibling (May 2024) built Google's reputation for aggressive price-performance. The 1.5 series carried Google's API lineup through 2024 via the -002 refresh before Gemini 2.0 opened the agentic era in early 2025. **Q: When was Gemini 1.5 Pro released?** A: Gemini 1.5 Pro was released by Google on Thursday, Feb 15 2024. **Q: Who made Gemini 1.5 Pro?** A: Gemini 1.5 Pro was built by Google. Builds the Gemini family of models through Google DeepMind. Integrates AI across Google products. **Q: Is Gemini 1.5 Pro open source?** A: No. Gemini 1.5 Pro is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Gemini 1.5 Pro?** A: Google's previous tracked release was Gemini 1.0 Ultra on Feb 8 2024, 7 days earlier. It was followed by Gemma 1 on Feb 21 2024. ### Gemini 1.0 Ultra - URL: https://aireleasetracker.com/model/google/gemini-1.0-ultra - Developer: Google - Release date: Thursday, Feb 8 2024 - Licensing: Proprietary - Benchmarks: none published in this dataset **Q: When was Gemini 1.0 Ultra released?** A: Gemini 1.0 Ultra was released by Google on Thursday, Feb 8 2024. **Q: Who made Gemini 1.0 Ultra?** A: Gemini 1.0 Ultra was built by Google. Builds the Gemini family of models through Google DeepMind. Integrates AI across Google products. **Q: Is Gemini 1.0 Ultra open source?** A: No. Gemini 1.0 Ultra is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Gemini 1.0 Ultra?** A: Google's previous tracked release was Gemini 1.0 Pro on Dec 13 2023, 57 days earlier. It was followed by Gemini 1.5 Pro on Feb 15 2024. ### DeepSeek-Math - URL: https://aireleasetracker.com/model/deepseek/deepseek-math - Developer: DeepSeek - Release date: Monday, Feb 5 2024 - Licensing: Open Weight - Benchmarks: none published in this dataset **Q: When was DeepSeek-Math released?** A: DeepSeek-Math was released by DeepSeek on Monday, Feb 5 2024. **Q: Who made DeepSeek-Math?** A: DeepSeek-Math was built by DeepSeek. Chinese AI lab known for efficient, open-weight models. Gained attention for strong performance at lower cost. **Q: Is DeepSeek-Math open source?** A: Partly. DeepSeek-Math is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after DeepSeek-Math?** A: DeepSeek's previous tracked release was DeepSeek-MoE on Jan 11 2024, 25 days earlier. It was followed by DeepSeek V2 on May 6 2024. ### Qwen1.5 - URL: https://aireleasetracker.com/model/qwen/qwen1.5 - Developer: Qwen - Release date: Sunday, Feb 4 2024 - Licensing: Open Weight - Parameters: 72B - Context window: 32k - Benchmarks: none published in this dataset **Q: When was Qwen1.5 released?** A: Qwen1.5 was released by Qwen on Sunday, Feb 4 2024. **Q: Who made Qwen1.5?** A: Qwen1.5 was built by Qwen. Alibaba's AI lab, building the Qwen family. The most prolific publisher of open-weight models of any major lab, alongside a proprietary Max and Plus tier sold through Alibaba Cloud. **Q: What is the context window of Qwen1.5?** A: Qwen1.5 has a context window of 32k. That is the maximum amount of input plus output the model can hold in a single request. **Q: How many parameters does Qwen1.5 have?** A: Qwen1.5 is reported at 72B parameters. **Q: Is Qwen1.5 open source?** A: Partly. Qwen1.5 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Qwen1.5?** A: Qwen's previous tracked release was Qwen-72B on Nov 30 2023, 66 days earlier. It was followed by Qwen2 on Jun 6 2024. ### Code Llama 70B - URL: https://aireleasetracker.com/model/meta/code-llama-70b - Developer: Meta - Release date: Monday, Jan 29 2024 - Licensing: Open Weight - Benchmarks: none published in this dataset **Q: When was Code Llama 70B released?** A: Code Llama 70B was released by Meta on Monday, Jan 29 2024. **Q: Who made Code Llama 70B?** A: Code Llama 70B was built by Meta. Develops the open-weight Llama series of models. Committed to open-source AI research. **Q: Is Code Llama 70B open source?** A: Partly. Code Llama 70B is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Code Llama 70B?** A: Meta's previous tracked release was Code Llama on Aug 24 2023, 158 days earlier. It was followed by LLaMA 3 (8B/70B) on Apr 18 2024. ### GLM-4 - URL: https://aireleasetracker.com/model/zai/glm-4 - Developer: Z.ai - Release date: Tuesday, Jan 16 2024 - Licensing: Proprietary - Benchmarks: none published in this dataset **Q: When was GLM-4 released?** A: GLM-4 was released by Z.ai on Tuesday, Jan 16 2024. **Q: Who made GLM-4?** A: GLM-4 was built by Z.ai. Chinese AI lab spun out of Tsinghua University (formerly Zhipu AI), building the open-weight GLM family. Rebranded internationally as Z.ai in 2025. **Q: Is GLM-4 open source?** A: No. GLM-4 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after GLM-4?** A: Z.ai's previous tracked release was ChatGLM-6B on Mar 14 2023, 308 days earlier. It was followed by GLM-4-9B on Jun 5 2024. ### DeepSeek-MoE - URL: https://aireleasetracker.com/model/deepseek/deepseek-moe - Developer: DeepSeek - Release date: Thursday, Jan 11 2024 - Licensing: Open Weight - Benchmarks: none published in this dataset **Q: When was DeepSeek-MoE released?** A: DeepSeek-MoE was released by DeepSeek on Thursday, Jan 11 2024. **Q: Who made DeepSeek-MoE?** A: DeepSeek-MoE was built by DeepSeek. Chinese AI lab known for efficient, open-weight models. Gained attention for strong performance at lower cost. **Q: Is DeepSeek-MoE open source?** A: Partly. DeepSeek-MoE is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after DeepSeek-MoE?** A: DeepSeek's previous tracked release was DeepSeek-LLM on Nov 29 2023, 43 days earlier. It was followed by DeepSeek-Math on Feb 5 2024. ### Gemini 1.0 Pro - URL: https://aireleasetracker.com/model/google/gemini-1.0-pro - Developer: Google - Release date: Wednesday, Dec 13 2023 - Licensing: Proprietary - Benchmarks: none published in this dataset **Q: When was Gemini 1.0 Pro released?** A: Gemini 1.0 Pro was released by Google on Wednesday, Dec 13 2023. **Q: Who made Gemini 1.0 Pro?** A: Gemini 1.0 Pro was built by Google. Builds the Gemini family of models through Google DeepMind. Integrates AI across Google products. **Q: Is Gemini 1.0 Pro open source?** A: No. Gemini 1.0 Pro is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Gemini 1.0 Pro?** A: Google's previous tracked release was Gemini 1.0 Nano on Dec 6 2023, 7 days earlier. It was followed by Gemini 1.0 Ultra on Feb 8 2024. ### Mistral Medium - URL: https://aireleasetracker.com/model/mistral/mistral-medium - Developer: Mistral - Release date: Monday, Dec 11 2023 - Licensing: Proprietary - Benchmarks: none published in this dataset **Q: When was Mistral Medium released?** A: Mistral Medium was released by Mistral on Monday, Dec 11 2023. **Q: Who made Mistral Medium?** A: Mistral Medium was built by Mistral. French AI company building open and commercial models. Founded in 2023 by former Meta and DeepMind researchers. **Q: Is Mistral Medium open source?** A: No. Mistral Medium is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Mistral Medium?** A: Mistral's previous tracked release was Mixtral 8×7B on Dec 8 2023, 3 days earlier. It was followed by Mistral Large on Feb 26 2024. ### Mixtral 8×7B - URL: https://aireleasetracker.com/model/mistral/mixtral-8×7b - Developer: Mistral - Release date: Friday, Dec 8 2023 - Licensing: Open Weight - Benchmarks: none published in this dataset Mixtral 8×7B, released December 8, 2023 — again as an unceremonious magnet link — brought sparse mixture-of-experts architecture to the open ecosystem. With eight expert networks of which two activate per token (~13B active of 47B total), it matched or beat GPT-3.5 and Llama 2 70B while inferring at the speed of a much smaller model, all under Apache 2.0. It was many developers' first hands-on proof that MoE was the future of efficient scaling, months before it became known that GPT-4 itself used the technique. Mixtral powered a generation of open chat products and was followed by the larger Mixtral 8×22B in April 2024. **Q: When was Mixtral 8×7B released?** A: Mixtral 8×7B was released by Mistral on Friday, Dec 8 2023. **Q: Who made Mixtral 8×7B?** A: Mixtral 8×7B was built by Mistral. French AI company building open and commercial models. Founded in 2023 by former Meta and DeepMind researchers. **Q: Is Mixtral 8×7B open source?** A: Partly. Mixtral 8×7B is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Mixtral 8×7B?** A: Mistral's previous tracked release was Mistral 7B on Sep 27 2023, 72 days earlier. It was followed by Mistral Medium on Dec 11 2023. ### Gemini 1.0 Nano - URL: https://aireleasetracker.com/model/google/gemini-1.0-nano - Developer: Google - Release date: Wednesday, Dec 6 2023 - Licensing: Proprietary - Benchmarks: none published in this dataset **Q: When was Gemini 1.0 Nano released?** A: Gemini 1.0 Nano was released by Google on Wednesday, Dec 6 2023. **Q: Who made Gemini 1.0 Nano?** A: Gemini 1.0 Nano was built by Google. Builds the Gemini family of models through Google DeepMind. Integrates AI across Google products. **Q: Is Gemini 1.0 Nano open source?** A: No. Gemini 1.0 Nano is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Gemini 1.0 Nano?** A: Google's previous tracked release was Bard on Mar 21 2023, 260 days earlier. It was followed by Gemini 1.0 Pro on Dec 13 2023. ### Qwen-72B - URL: https://aireleasetracker.com/model/qwen/qwen-72b - Developer: Qwen - Release date: Thursday, Nov 30 2023 - Licensing: Open Weight - Parameters: 72B - Context window: 32k - Benchmarks: none published in this dataset **Q: When was Qwen-72B released?** A: Qwen-72B was released by Qwen on Thursday, Nov 30 2023. **Q: Who made Qwen-72B?** A: Qwen-72B was built by Qwen. Alibaba's AI lab, building the Qwen family. The most prolific publisher of open-weight models of any major lab, alongside a proprietary Max and Plus tier sold through Alibaba Cloud. **Q: What is the context window of Qwen-72B?** A: Qwen-72B has a context window of 32k. That is the maximum amount of input plus output the model can hold in a single request. **Q: How many parameters does Qwen-72B have?** A: Qwen-72B is reported at 72B parameters. **Q: Is Qwen-72B open source?** A: Partly. Qwen-72B is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Qwen-72B?** A: Qwen's previous tracked release was Qwen-7B on Aug 3 2023, 119 days earlier. It was followed by Qwen1.5 on Feb 4 2024. ### DeepSeek-LLM - URL: https://aireleasetracker.com/model/deepseek/deepseek-llm - Developer: DeepSeek - Release date: Wednesday, Nov 29 2023 - Licensing: Open Weight - Benchmarks: none published in this dataset **Q: When was DeepSeek-LLM released?** A: DeepSeek-LLM was released by DeepSeek on Wednesday, Nov 29 2023. **Q: Who made DeepSeek-LLM?** A: DeepSeek-LLM was built by DeepSeek. Chinese AI lab known for efficient, open-weight models. Gained attention for strong performance at lower cost. **Q: Is DeepSeek-LLM open source?** A: Partly. DeepSeek-LLM is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after DeepSeek-LLM?** A: DeepSeek's previous tracked release was DeepSeek Coder on Nov 2 2023, 27 days earlier. It was followed by DeepSeek-MoE on Jan 11 2024. ### Claude 2.1 - URL: https://aireleasetracker.com/model/anthropic/claude-2.1 - Developer: Anthropic - Release date: Tuesday, Nov 21 2023 - Licensing: Proprietary - Context window: 200k (about 150,000 words) - Benchmarks: none published in this dataset Claude 2.1, released November 21, 2023, doubled the context window to 200,000 tokens — about 150,000 words, or more than 500 pages — the largest of any commercially available model at the time. It landed in the same week that OpenAI was consumed by the firing and reinstatement of Sam Altman, and Anthropic used the opening to push long-document work as the thing Claude was for: contract review, financial filings, entire technical specifications in a single prompt. Anthropic published no new academic benchmark scores with the release, leading instead on reliability: a 2x reduction in false statements compared with Claude 2.0, and a 30% drop in incorrect answers on document-comprehension tasks. The release also brought system prompts and a beta of tool use. Independent needle-in-a-haystack testing published the day after launch found recall degrading once prompts ran past roughly 90K tokens, and varying with where in the document the buried fact sat — an early public demonstration that having a large context window and using all of it reliably are two different problems. Anthropic followed up two weeks later with prompting guidance that recovered much of the gap. **Q: When was Claude 2.1 released?** A: Claude 2.1 was released by Anthropic on Tuesday, Nov 21 2023. **Q: Who made Claude 2.1?** A: Claude 2.1 was built by Anthropic. AI safety company building the Claude family of models. Founded in 2021 by former OpenAI researchers. **Q: What is the context window of Claude 2.1?** A: Claude 2.1 has a context window of 200k — roughly 150,000 words of text. That is the maximum amount of input plus output the model can hold in a single request. **Q: Is Claude 2.1 open source?** A: No. Claude 2.1 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Claude 2.1?** A: Anthropic's previous tracked release was Claude Instant 1.2 on Aug 9 2023, 104 days earlier. It was followed by Claude 3 Haiku on Mar 4 2024. ### GPT-4 Turbo - URL: https://aireleasetracker.com/model/openai/gpt-4-turbo - Developer: OpenAI - Release date: Monday, Nov 6 2023 - Licensing: Proprietary - Context window: 128k - Benchmarks published at release: - GPQA Diamond: 42.5% **Q: When was GPT-4 Turbo released?** A: GPT-4 Turbo was released by OpenAI on Monday, Nov 6 2023. **Q: Who made GPT-4 Turbo?** A: GPT-4 Turbo was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: What benchmark scores did GPT-4 Turbo get?** A: GPT-4 Turbo reports 1 tracked benchmark score — GPQA Diamond: 42.5%. Scores are the figures published at release by OpenAI. **Q: What is the context window of GPT-4 Turbo?** A: GPT-4 Turbo has a context window of 128k. That is the maximum amount of input plus output the model can hold in a single request. **Q: Is GPT-4 Turbo open source?** A: No. GPT-4 Turbo is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after GPT-4 Turbo?** A: OpenAI's previous tracked release was GPT-4 on Mar 14 2023, 237 days earlier. It was followed by GPT-4o on May 13 2024. ### Grok‑1 - URL: https://aireleasetracker.com/model/xai/grok‑1 - Developer: SpaceXAI - Release date: Friday, Nov 3 2023 - Licensing: Open Weight - Benchmarks: none published in this dataset **Q: When was Grok‑1 released?** A: Grok‑1 was released by SpaceXAI on Friday, Nov 3 2023. **Q: Who made Grok‑1?** A: Grok‑1 was built by SpaceXAI. Elon Musk's AI company building the Grok series of models. Founded in 2023 as xAI, now part of SpaceX. **Q: Is Grok‑1 open source?** A: Partly. Grok‑1 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Grok‑1?** A: It was followed by Grok‑1.5 on Mar 29 2024. ### DeepSeek Coder - URL: https://aireleasetracker.com/model/deepseek/deepseek-coder - Developer: DeepSeek - Release date: Thursday, Nov 2 2023 - Licensing: Open Weight - Benchmarks: none published in this dataset **Q: When was DeepSeek Coder released?** A: DeepSeek Coder was released by DeepSeek on Thursday, Nov 2 2023. **Q: Who made DeepSeek Coder?** A: DeepSeek Coder was built by DeepSeek. Chinese AI lab known for efficient, open-weight models. Gained attention for strong performance at lower cost. **Q: Is DeepSeek Coder open source?** A: Partly. DeepSeek Coder is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after DeepSeek Coder?** A: It was followed by DeepSeek-LLM on Nov 29 2023. ### Kimi Chat - URL: https://aireleasetracker.com/model/moonshot/kimi-chat - Developer: Moonshot AI - Release date: Monday, Oct 9 2023 - Licensing: Proprietary - Benchmarks: none published in this dataset **Q: When was Kimi Chat released?** A: Kimi Chat was released by Moonshot AI on Monday, Oct 9 2023. **Q: Who made Kimi Chat?** A: Kimi Chat was built by Moonshot AI. Chinese AI company building the Kimi family of models. Founded in 2023, known for long-context and efficient MoE architectures. **Q: Is Kimi Chat open source?** A: No. Kimi Chat is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Kimi Chat?** A: It was followed by Kimi K1.5 on Jan 20 2025. ### Mistral 7B - URL: https://aireleasetracker.com/model/mistral/mistral-7b - Developer: Mistral - Release date: Wednesday, Sep 27 2023 - Licensing: Open Weight - Benchmarks: none published in this dataset Mistral 7B, released September 27, 2023 as a bare magnet link on X, was the debut model of the French startup and an instant classic: at 7 billion parameters it outperformed Llama 2 13B across benchmarks, using grouped-query and sliding-window attention to squeeze unprecedented quality from a small dense model. The Apache 2.0 license made it usable anywhere, by anyone. It became one of the most fine-tuned and deployed open models ever — the default base for local assistants, edge deployments, and countless derivatives — and it established Mistral, then a months-old company, as Europe's frontier lab. The mixture-of-experts Mixtral 8×7B followed just ten weeks later. **Q: When was Mistral 7B released?** A: Mistral 7B was released by Mistral on Wednesday, Sep 27 2023. **Q: Who made Mistral 7B?** A: Mistral 7B was built by Mistral. French AI company building open and commercial models. Founded in 2023 by former Meta and DeepMind researchers. **Q: Is Mistral 7B open source?** A: Partly. Mistral 7B is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Mistral 7B?** A: It was followed by Mixtral 8×7B on Dec 8 2023. ### Code Llama - URL: https://aireleasetracker.com/model/meta/code-llama - Developer: Meta - Release date: Thursday, Aug 24 2023 - Licensing: Open Weight - Benchmarks: none published in this dataset **Q: When was Code Llama released?** A: Code Llama was released by Meta on Thursday, Aug 24 2023. **Q: Who made Code Llama?** A: Code Llama was built by Meta. Develops the open-weight Llama series of models. Committed to open-source AI research. **Q: Is Code Llama open source?** A: Partly. Code Llama is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Code Llama?** A: Meta's previous tracked release was LLaMA 2 on Jul 18 2023, 37 days earlier. It was followed by Code Llama 70B on Jan 29 2024. ### Claude Instant 1.2 - URL: https://aireleasetracker.com/model/anthropic/claude-instant-1.2 - Developer: Anthropic - Release date: Wednesday, Aug 9 2023 - Licensing: Proprietary - Context window: 100k - Benchmarks published at release: - HumanEval: 58.7% - GSM8K: 86.7% **Q: When was Claude Instant 1.2 released?** A: Claude Instant 1.2 was released by Anthropic on Wednesday, Aug 9 2023. **Q: Who made Claude Instant 1.2?** A: Claude Instant 1.2 was built by Anthropic. AI safety company building the Claude family of models. Founded in 2021 by former OpenAI researchers. **Q: What benchmark scores did Claude Instant 1.2 get?** A: Claude Instant 1.2 reports 2 tracked benchmark scores — HumanEval: 58.7%; GSM8K: 86.7%. Scores are the figures published at release by Anthropic. **Q: What is the context window of Claude Instant 1.2?** A: Claude Instant 1.2 has a context window of 100k. That is the maximum amount of input plus output the model can hold in a single request. **Q: Is Claude Instant 1.2 open source?** A: No. Claude Instant 1.2 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Claude Instant 1.2?** A: Anthropic's previous tracked release was Claude 2 on Jul 11 2023, 29 days earlier. It was followed by Claude 2.1 on Nov 21 2023. ### Qwen-7B - URL: https://aireleasetracker.com/model/qwen/qwen-7b - Developer: Qwen - Release date: Thursday, Aug 3 2023 - Licensing: Open Weight - Parameters: 7B - Context window: 2k - Benchmarks: none published in this dataset Qwen-7B, released August 3, 2023, was the first Qwen model Alibaba put on Hugging Face, and the point at which Tongyi Qianwen stopped being only a hosted chatbot. Alibaba had unveiled Tongyi Qianwen that April as an internal-facing product, but it could not be opened to the Chinese public until regulatory clearance came through in September. Publishing the 7B weights first turned a product announcement into something researchers outside Alibaba could actually build on. The architecture drew heavily on Meta's Llama, and the release carried a bespoke Tongyi Qianwen licence rather than a standard open one — a pattern Qwen kept for its largest models for years afterward. A 14B model followed in September 2023, then Qwen-72B and Qwen-1.8B on November 30, 2023, which gave the family a genuine flagship and established the size-ladder strategy that every later generation repeated. **Q: When was Qwen-7B released?** A: Qwen-7B was released by Qwen on Thursday, Aug 3 2023. **Q: Who made Qwen-7B?** A: Qwen-7B was built by Qwen. Alibaba's AI lab, building the Qwen family. The most prolific publisher of open-weight models of any major lab, alongside a proprietary Max and Plus tier sold through Alibaba Cloud. **Q: What is the context window of Qwen-7B?** A: Qwen-7B has a context window of 2k. That is the maximum amount of input plus output the model can hold in a single request. **Q: How many parameters does Qwen-7B have?** A: Qwen-7B is reported at 7B parameters. **Q: Is Qwen-7B open source?** A: Partly. Qwen-7B is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after Qwen-7B?** A: It was followed by Qwen-72B on Nov 30 2023. ### LLaMA 2 - URL: https://aireleasetracker.com/model/meta/llama-2 - Developer: Meta - Release date: Tuesday, Jul 18 2023 - Licensing: Open Weight - Benchmarks: none published in this dataset Llama 2, released July 18, 2023 in 7B, 13B, and 70B sizes, was the release that created the open-weight ecosystem. Where the original LLaMA leaked under a research-only license, Llama 2 was free for commercial use and distributed openly in partnership with Microsoft Azure — a direct strategic bet against the closed-API model of OpenAI and Anthropic. Within months it was the foundation of thousands of fine-tunes, the standard base model for academic research, and the reason tools like llama.cpp and vLLM exist in their modern form. Code Llama followed in August 2023, and the line's successor Llama 3 arrived in April 2024. **Q: When was LLaMA 2 released?** A: LLaMA 2 was released by Meta on Tuesday, Jul 18 2023. **Q: Who made LLaMA 2?** A: LLaMA 2 was built by Meta. Develops the open-weight Llama series of models. Committed to open-source AI research. **Q: Is LLaMA 2 open source?** A: Partly. LLaMA 2 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after LLaMA 2?** A: Meta's previous tracked release was LLaMA 1 on Feb 24 2023, 144 days earlier. It was followed by Code Llama on Aug 24 2023. ### Claude 2 - URL: https://aireleasetracker.com/model/anthropic/claude-2 - Developer: Anthropic - Release date: Tuesday, Jul 11 2023 - Licensing: Proprietary - Context window: 100k (about 75,000 words) - Benchmarks published at release: - HumanEval: 71.2% - MMLU: 78.5% - GSM8K: 88% Claude 2, released July 11, 2023, was the first Claude anyone could simply sign up and use. It launched alongside claude.ai as a public beta in the US and UK, ending the waitlist-and-API-key period that had kept Claude 1 to a narrow developer audience. It carried a 100K-token context window — roughly 75,000 words — which at the time was the largest available from any major AI lab and made whole-book and whole-codebase prompting a mainstream idea rather than a research demo. Anthropic pitched it on measured gains over Claude 1.3: 71.2% on Codex HumanEval, up from 56.0%; 88.0% on GSM8K; 78.5% on MMLU; and 76.5% on the multistate bar exam. Those were respectable rather than record-setting figures next to GPT-4, and Claude 2's real significance was distribution — it put Anthropic in front of consumers four months after GPT-4 shipped, and set up the context-window expansion that arrived with Claude 2.1 in November. **Q: When was Claude 2 released?** A: Claude 2 was released by Anthropic on Tuesday, Jul 11 2023. **Q: Who made Claude 2?** A: Claude 2 was built by Anthropic. AI safety company building the Claude family of models. Founded in 2021 by former OpenAI researchers. **Q: What benchmark scores did Claude 2 get?** A: Claude 2 reports 3 tracked benchmark scores — HumanEval: 71.2%; MMLU: 78.5%; GSM8K: 88%. Scores are the figures published at release by Anthropic. **Q: What is the context window of Claude 2?** A: Claude 2 has a context window of 100k — roughly 75,000 words of text. That is the maximum amount of input plus output the model can hold in a single request. **Q: Is Claude 2 open source?** A: No. Claude 2 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Claude 2?** A: Anthropic's previous tracked release was Claude 1 on Mar 14 2023, 119 days earlier. It was followed by Claude Instant 1.2 on Aug 9 2023. ### Bard - URL: https://aireleasetracker.com/model/google/bard - Developer: Google - Release date: Tuesday, Mar 21 2023 - Licensing: Proprietary - Benchmarks: none published in this dataset **Q: When was Bard released?** A: Bard was released by Google on Tuesday, Mar 21 2023. **Q: Who made Bard?** A: Bard was built by Google. Builds the Gemini family of models through Google DeepMind. Integrates AI across Google products. **Q: Is Bard open source?** A: No. Bard is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Bard?** A: It was followed by Gemini 1.0 Nano on Dec 6 2023. ### Claude 1 - URL: https://aireleasetracker.com/model/anthropic/claude-1 - Developer: Anthropic - Release date: Tuesday, Mar 14 2023 - Licensing: Proprietary - Context window: 9k - Benchmarks: none published in this dataset **Q: When was Claude 1 released?** A: Claude 1 was released by Anthropic on Tuesday, Mar 14 2023. **Q: Who made Claude 1?** A: Claude 1 was built by Anthropic. AI safety company building the Claude family of models. Founded in 2021 by former OpenAI researchers. **Q: What is the context window of Claude 1?** A: Claude 1 has a context window of 9k. That is the maximum amount of input plus output the model can hold in a single request. **Q: Is Claude 1 open source?** A: No. Claude 1 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after Claude 1?** A: It was followed by Claude 2 on Jul 11 2023. ### GPT-4 - URL: https://aireleasetracker.com/model/openai/gpt-4 - Developer: OpenAI - Release date: Tuesday, Mar 14 2023 - Licensing: Proprietary - Context window: 8k - Benchmarks published at release: - HumanEval: 67% - MMLU: 86.4% - GSM8K: 92% GPT-4, released March 14, 2023, was the model that proved scaling kept working. It passed a simulated bar exam in the top 10% of test takers, accepted image input, and was so far ahead of contemporaries that it anchored ChatGPT Plus, Microsoft's Bing Chat, and the first wave of serious commercial LLM applications. Notably, Anthropic released Claude 1 the very same day. OpenAI published almost no technical details — no parameter count, architecture, or training data — a break from research-lab tradition that set the industry's closed-frontier norm. GPT-4 headed OpenAI's lineup for more than a year, extended by GPT-4 Turbo's 128K context window in November 2023, until GPT-4o replaced it in May 2024. **Q: When was GPT-4 released?** A: GPT-4 was released by OpenAI on Tuesday, Mar 14 2023. **Q: Who made GPT-4?** A: GPT-4 was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: What benchmark scores did GPT-4 get?** A: GPT-4 reports 3 tracked benchmark scores — HumanEval: 67%; MMLU: 86.4%; GSM8K: 92%. Scores are the figures published at release by OpenAI. **Q: What is the context window of GPT-4?** A: GPT-4 has a context window of 8k. That is the maximum amount of input plus output the model can hold in a single request. **Q: Is GPT-4 open source?** A: No. GPT-4 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after GPT-4?** A: OpenAI's previous tracked release was GPT-3.5 on Nov 30 2022, 104 days earlier. It was followed by GPT-4 Turbo on Nov 6 2023. ### ChatGLM-6B - URL: https://aireleasetracker.com/model/zai/chatglm-6b - Developer: Z.ai - Release date: Tuesday, Mar 14 2023 - Licensing: Open Weight - Parameters: 6B - Benchmarks: none published in this dataset **Q: When was ChatGLM-6B released?** A: ChatGLM-6B was released by Z.ai on Tuesday, Mar 14 2023. **Q: Who made ChatGLM-6B?** A: ChatGLM-6B was built by Z.ai. Chinese AI lab spun out of Tsinghua University (formerly Zhipu AI), building the open-weight GLM family. Rebranded internationally as Z.ai in 2025. **Q: How many parameters does ChatGLM-6B have?** A: ChatGLM-6B is reported at 6B parameters. **Q: Is ChatGLM-6B open source?** A: Partly. ChatGLM-6B is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after ChatGLM-6B?** A: It was followed by GLM-4 on Jan 16 2024. ### LLaMA 1 - URL: https://aireleasetracker.com/model/meta/llama-1 - Developer: Meta - Release date: Friday, Feb 24 2023 - Licensing: Open Weight - Benchmarks: none published in this dataset **Q: When was LLaMA 1 released?** A: LLaMA 1 was released by Meta on Friday, Feb 24 2023. **Q: Who made LLaMA 1?** A: LLaMA 1 was built by Meta. Develops the open-weight Llama series of models. Committed to open-source AI research. **Q: Is LLaMA 1 open source?** A: Partly. LLaMA 1 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense. **Q: What came before and after LLaMA 1?** A: It was followed by LLaMA 2 on Jul 18 2023. ### GPT-3.5 - URL: https://aireleasetracker.com/model/openai/gpt-3.5 - Developer: OpenAI - Release date: Wednesday, Nov 30 2022 - Licensing: Proprietary - Benchmarks: none published in this dataset GPT-3.5 is the model that started the modern AI race. On November 30, 2022 OpenAI wrapped it in a simple chat interface called ChatGPT, which reached 100 million users within two months — the fastest-growing consumer application in history at the time. Every release on this tracker's timeline is measured from that day. Technically a fine-tuned descendant of GPT-3 trained with reinforcement learning from human feedback, GPT-3.5 was quickly surpassed — GPT-4 arrived less than four months later — but its free availability kept it the world's most-used language model well into 2024. OpenAI never published a benchmark table for GPT-3.5 on its own. The figures on record come from the GPT-4 technical report in March 2023, where GPT-3.5 was the baseline the new model was announced against: 70.0% on MMLU, 57.1% on GSM8K and 48.1% on HumanEval. Those were strong numbers for the moment and are useful now mainly as a starting line — every one of them was a solved problem within two years. **Q: When was GPT-3.5 released?** A: GPT-3.5 was released by OpenAI on Wednesday, Nov 30 2022. **Q: Who made GPT-3.5?** A: GPT-3.5 was built by OpenAI. Creators of ChatGPT and the GPT series of models. Pioneered large-scale language model research. **Q: Is GPT-3.5 open source?** A: No. GPT-3.5 is a proprietary model. The weights are not published — it is available only through the provider's own API, apps, or partner platforms. **Q: What came before and after GPT-3.5?** A: It was followed by GPT-4 on Mar 14 2023.