Compare AI models
| Specifications | ||
ParametersThe learned values that shape how a model responds. B means billion and T means trillion. For mixture-of-experts models, the total count is larger than the number active for each token. Parameter count alone does not measure quality. | 4B | — |
Context windowHow much information the model can work with in one request, measured in tokens — pieces of words or other data. This includes your question, attachments, conversation and reply. Limits can vary by model variant and provider. | 256k | 1.05M |
| API pricingUSD per 1M tokens · lower wins · base tier | ||
Input priceWhat you pay for everything you send the model — your question, plus any documents or earlier conversation you include with it. | — | $0.20 |
Output priceWhat you pay for the text the model writes back. It is normally the dearer half: producing an answer costs more than reading one. | — | $1.20 |
Cached input priceA reduced rate for text you send over and over. If every request starts with the same instructions or the same document, the provider keeps a copy ready and charges less to read it again. | — | $0.02 |
Cheapest inputLowest input rate across every provider on OpenRouter, the lab included. The cheapest endpoint may run a quantised build or a shorter context — see "Available from" on the model page. | — | $0.10OpenAI |
Cheapest outputLowest output rate across every provider on OpenRouter, the lab included. May come from a different provider than the cheapest input. | — | $0.60OpenAI |
| — | $100OpenAI list price | |
| Benchmarks | ||
Terminal-Bench 2.1Agentic terminal coding — Can the AI work in a command-line terminal — running commands and finishing technical setup tasks the way a developer would? Higher is better. | 18.1% | 84.7% |
GPQA DiamondScience — Graduate-level science questions in biology, physics, and chemistry — hard enough that subject-matter PhDs score around 65%. Higher is better. | 72.41% | 92.3% |
BullshitBench v2Nonsense detection — 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. | — | 41% |
Gray Swan IPI · k = 1Prompt injection robustness — 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. | — | 8.3% |
Gray Swan IPI · k = 10Prompt injection robustness — 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. | — | 38.6% |
Gray Swan IPI · k = 15Prompt injection robustness — 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. | — | 43.9% |
SWE-Bench ProAgentic coding — 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. | — | 62.7% |
SWE-Bench VerifiedCoding — 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. | 59.5% | — |
CursorBench 4.0Agentic coding — Cursor's own test of coding agents on ambiguous, multi-file tasks taken from real Cursor sessions — editing, refactoring, investigating a codebase, understanding what the user meant, managing jobs and following a design. Cursor runs each model at several reasoning efforts; each release here carries the score of its best listed effort. Scores aren't comparable with earlier CursorBench versions. Higher is better. | — | 35.9% |
DeepSWE 1.1Agentic coding — 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. | — | 67.2% |
Supabase Evals · with skillsSupabase coding — 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. | — | 63.8% |
Supabase Evals · no skillsSupabase coding — 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. | — | 42% |
BenchCAD · without toolsCAD programming — geometric overlap of generated CAD parts without tool access | — | 63.1% |
BenchCAD · with toolsCAD programming — geometric overlap of generated CAD parts with tool access | — | 73.9% |
Frontier-Bench v0.1Agentic computer work — 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. | — | 14.3% |
Terminal-Bench 4.0Agentic terminal coding — Can the AI work in a command-line terminal — running commands and finishing technical setup tasks the way a developer would? Version 4.0 recalibrated how much time, CPU and memory each task gets, removed eight tasks and fixed nineteen, so fewer runs fail for reasons that have nothing to do with the model. Scores are not comparable with earlier versions. Higher is better. | — | 17.27% |
ToolathlonGeneral tool use — Tests how well the AI uses everyday real-world tools and apps to get things done. Higher is better. | — | 53.4% |
BrowseCompWeb browsing — Can the AI browse the web and track down hard-to-find answers? Higher is better. | — | 83.3% |
ARC-AGI-3Novel problem-solving — 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. | — | 0.18% |
ARC-AGI-2Abstract reasoning — Puzzle-style tests of abstract reasoning and pattern-finding — the kind of thing people find easy but AIs often struggle with. Higher is better. | — | 59.54% |
FrontierMath · Tier 4 (v2)Advanced math — The rebuilt edition of FrontierMath's hardest tier — research-level maths of the kind professional mathematicians work on. It is a different question set from the first Tier 4, and scores on it run far higher, so read the two as separate tests rather than progress. Higher is better. | — | 58.5% |
GeneBench-ProBiology — Real genomics and biomedical analyses done end to end: the AI gets a messy dataset and a question, and has to work through the chain of statistical decisions to a verifiable answer — the job a computational biologist does before a research or clinical decision gets made. Higher is better. | — | 10.8% |
MedChemBench (Internal)Chemistry — OpenAI's own drug-discovery test: reading chemical structures, predicting how potent or toxic a compound will be, choosing between candidate molecules, and planning a synthesis route — the everyday judgement calls of a medicinal chemist. Higher is better. | — | 30.4% |
LifeSciBenchReasoning — performance on research tasks graded by life sciences experts | — | 51.2% |
FrontierMath · Tiers 1–3 (v2)Reasoning — research mathematics on the rebuilt second edition of tiers 1–3 | — | 78.6% |
IFBenchInstruction following — Tests whether the AI can follow detailed instructions and satisfy multiple constraints at once. Higher is better. | 77.94% | — |
Agent's Last Exam · pass@1Agentic computer use — 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. | — | 50.3% |
AutomationBenchBusiness workflows — 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. | — | 14.9% |
HealthBench Professional · OpenAI grading, length-adjustedHealth — official OpenAI grading with the response-length penalty applied | — | 55.7% |
GDP.PDFDocument comprehension — 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. | — | 22.7% |
MMMU-Pro · with toolsMultimodal reasoning — multimodal understanding and reasoning with access to tools | — | 79.5% |
MMMU-ProMultimodal reasoning — A tougher version of MMMU — college-level questions that mix images, diagrams, and text together. Higher is better. | — | 78.4% |
MRCR v2 (8-needle) · 256K–512KLong context — mean match ratio for eight needles within the indicated context band | — | 41.3% |
MRCR v2 (8-needle) · 512K–1MLong context — mean match ratio for eight needles within the indicated context band | — | 41.3% |
threejsevalCommunity preference (Three.js) — Every model gets the same prompt — "the Eiffel Tower", "a glass fishbowl", "a robot arm picking toys into a box" — and builds a 3D scene in Three.js. Real people then see two scenes side by side, names hidden, and vote for the one they prefer. The votes become a chess-style Elo rating on threejseval.com, averaged across all the prompts. It measures whether the scene looks and moves right to a human eye, not whether the code passes a test. Higher is better. | — | 1270 |
| Overview | ||
| Company | NVIDIA | OpenAI |
| Release date | Oct 16 2026 | Jun 26 2026 |
| Access | Open Weight | Closed |
| Model details | View model | View model |
Other comparisons
Nemotron Labs Dory 1.0 4BvsClaude Fable 5.2GPT-5.6 LunavsClaude Fable 5.2Nemotron Labs Dory 1.0 4BvsGemini 4 CarbonGPT-5.6 LunavsGemini 4 CarbonNemotron Labs Dory 1.0 4BvsMuse EmberGPT-5.6 LunavsMuse EmberNemotron Labs Dory 1.0 4BvsGrok 4.7GPT-5.6 LunavsGrok 4.7Nemotron Labs Dory 1.0 4BvsDeepSeek-V4.1-Flash-1017GPT-5.6 LunavsDeepSeek-V4.1-Flash-1017Nemotron Labs Dory 1.0 4BvsMistral Large 4GPT-5.6 LunavsMistral Large 4Frequently asked questions
GPT-5.6 Luna leads Nemotron Labs Dory 1.0 4B on 2 of the 2 benchmarks they both report (Terminal-Bench 2.1, GPQA Diamond). Only GPT-5.6 Luna has a verified first-party price for an available API: $0.20 per million input tokens and $1.20 per million output tokens. No currently available pay-as-you-go API rate is tracked for Nemotron Labs Dory 1.0 4B. GPT-5.6 Luna shipped 112 days before Nemotron Labs Dory 1.0 4B, so benchmark comparisons should account for the intervening progress.
Context windows are 256k (Nemotron Labs Dory 1.0 4B) vs 1.05M (GPT-5.6 Luna). Nemotron Labs Dory 1.0 4B is open weight, while GPT-5.6 Luna is closed.
On Terminal-Bench 2.1, GPT-5.6 Luna leads at 84.7% vs Nemotron Labs Dory 1.0 4B at 18.1%. On GPQA Diamond, GPT-5.6 Luna leads at 92.3% vs Nemotron Labs Dory 1.0 4B at 72.41%.