AI Model Release Tracker

GPT-6 Astra

Released

GPT-6 Astra is an AI model released by OpenAI on Thursday, Sep 3 2026, 24 days after GPT-5.6-Cyber. It has a 1.05M token context window. Benchmark results (shown below) cover Auto-review circumvention (Internal), FrontierCode v1.1 (Main), AA Coding Agent Index, Database Migration Tasks (OpenAI Internal), BenchCAD, Terminal-Bench-Science 0.1, and 22 more.

API pricing

Up to 272K input tokens
InputWhat you pay for everything you send the model — your question, plus any documents or earlier conversation you include with it.
$10.00
Cached inputA 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.
$1.00
Cache writeA one-off charge for storing text so later requests can reuse it at the cheaper cached rate.
$12.50
OutputWhat you pay for the text the model writes back. It is normally the dearer half: producing an answer costs more than reading one.
$50.00
Over 272K input tokens
InputWhat you pay for everything you send the model — your question, plus any documents or earlier conversation you include with it.
$20.00
Cached inputA 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.
$2.00
Cache writeA one-off charge for storing text so later requests can reuse it at the cheaper cached rate.
$25.00
OutputWhat you pay for the text the model writes back. It is normally the dearer half: producing an answer costs more than reading one.
$75.00
  • Reasoning or thinking is supported.
  • OpenAI also publishes a Fast mode for this model at exactly twice the standard rate in every column — $20.00 per million input tokens and $100.00 per million output at short context.
USD per 1M tokensEvery price here is for one million tokens. A token is roughly three-quarters of a word, so a million tokens is about 750,000 words of text.Model mappingVerified September 3, 2026Read from developers.openai.com. Click the check to open the source.

Available from

OpenAI—flex$5.00$25.001M
OpenAI——$10.00$50.001M
Azure——$10.00$50.001M
Azureus—$11.00$55.001M
Amazon Bedrockus-west-2—$11.00$55.001M
OpenAI—fast$20.00$100.001M
OpenAI—ultrafast$60.00$300.001M
USD per 1M tokensEvery price here is for one million tokens. A token is roughly three-quarters of a word, so a million tokens is about 750,000 words of text.Checked October 18, 2026Third-party rates from openrouter.ai, not the lab’s own list price.

Benchmarks

Coding

FrontierCode v1.1 (Main)Agentic coding — 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.
53.3%main split
64.5%extended split
main split
#2 of 8
extended split
#1
AA Coding Agent IndexAgentic coding — Artificial Analysis' overall score for coding agents, combining three coding benchmarks with what each run costs and how many tokens it burns. It rates a model paired with a particular agent harness rather than the model alone, so the same model scores differently in different tools. Higher is better.
67
#1
Database Migration Tasks (OpenAI Internal)Database migrations — OpenAI's own test of moving a database from one schema or system to another without breaking what depends on it — the migration work that has to be right the first time. Higher is better.
63.9%
BenchCADCAD programming — Can the AI do mechanical design as code? Given a drawing or a description of an industrial part — a gear, a spring, a drill bit — it has to write or edit the parametric CAD program that builds it, and the program is run to check the shape really comes out right. Higher is better.
95.9%
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.
74.1%
#8 of 36
Next.js EvalsNext.js coding — 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.
90%
#7 of 33

Terminal & CLI

Terminal-Bench-Science 0.1Agentic scientific computing — The same command-line setup as Terminal-Bench, pointed at scientific work: the AI has to drive research tooling and computational workflows through to a result, rather than administer a machine. Version 0.1 is the first release of the task set, and scores run lower than on the general board. Higher is better.
64.6%
#1
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.
57.9%
#5 of 29

Agentic & tool use

ExploitBenchCybersecurity — 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.
100%
#1
Agent's Last ExamAgentic 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.
59.3%pass@1
#1
SRE-BenchSite reliability — Can the AI keep production running? It is dropped into a broken Kubernetes system and has to diagnose the incident and fix it safely, the way an on-call site-reliability engineer would. Scored here on the best of four attempts. Higher is better.
99.2%
BrowseCompWeb browsing — Can the AI browse the web and track down hard-to-find answers? Higher is better.
91.5%
#2 of 30
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.
41.4%
#10 of 19

Reasoning & science

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.
99.9%
#1
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.via BenchLM
95%
#1
FrontierMathAdvanced 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.
97.6%Tier 4 (v2)
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.
39%
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.
49.7%
GPQA DiamondScience — Graduate-level science questions in biology, physics, and chemistry — hard enough that subject-matter PhDs score around 65%. Higher is better.
96%
#1

Knowledge work

AA Intelligence IndexOverall intelligence — Artificial Analysis composite intelligence index across evals
61.2
#1
Design Tasks (OpenAI Internal)Design work — OpenAI's own set of professional design briefs, scored on whether the finished work is what a designer would have handed over. Higher is better.
50%
Data Science Tasks (OpenAI Internal)Data science — OpenAI's own set of data-science jobs — taking a dataset and a question through cleaning, analysis and a defensible answer. Higher is better.
40.9%

Healthcare

HealthBench ProfessionalHealth — The same physician-graded health conversations as HealthBench Professional, scored with an adjustment for how long the answer is — so a model cannot gain by padding its reply. The adjustment moves scores by several points, so these numbers are not interchangeable with the unadjusted ones. Higher is better.
63.4%length-adjusted

Multimodal

OpenScore String QuartetsSheet music — Can the AI read sheet music? It is shown scanned pages of string quartets and has to transcribe the notation — pitches, beams, accidentals and all — with the score measuring how close the transcription lands to the real thing. Runs 0 to 1, and higher is better.
0.84

Robustness

Auto-review circumvention (Internal)Safety-review circumvention — OpenAI's internal safety check on how often a model finds ways around its own automated review — the guardrail that inspects what it is about to do. This one counts failures, so lower is better and zero is the goal.
0%
#1
CheatBenchCheating rate — Measures how often AI agents try to cheat on difficult assignments, such as reading hidden answers, copying work or manipulating grading. The overall score gives equal weight to ten categories; the sycophancy category measures how far an agent shifts its beliefs toward a user's stated views. Scores describe each model in its tested agent setup, and task success is measured separately. Lower is better.via CheatBench
47.4%
#5 of 13
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.
69%
#18 of 97

Community preference

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.
1886
#3 of 27

Source: CheatBench, retrieved 25 September 2026 · Source: BenchLM, retrieved 7 September 2026. Other sources are identified on the linked benchmark pages.

Compare GPT-6 Astra with

GPT-6 Astra

Suggested comparisons

GPT-6 AstravsGPT-6.1 SolGPT-6 AstravsClaude Haiku 5.5GPT-6 AstravsGemini 4 ArgonGPT-6 AstravsMuse Spark OpenGPT-6 AstravsGrok 4.7GPT-6 AstravsDeepSeek-V4.1-Flash-1017GPT-6 AstravsMistral Large 4GPT-6 AstravsKimi K3GPT-6 AstravsGLM-5.3-FlashGPT-6 AstravsQwen3.8-Max-0902GPT-6 AstravsNemotron 3.5 Super VL

About

GPT-6 Astra, released September 3, 2026, opened OpenAI's GPT-6 generation and was the company's first new general-purpose flagship since the three-model GPT-5.6 family — Sol, Terra and Luna — arrived on June 26, 2026. OpenAI had trailed the model for a month before launch without ever announcing it as a product: on August 1, 2026 it named Astra "our next major model" in a research post, reporting that an internal version had produced new results on ten long-standing open mathematical problems, each with a machine-checkable proof in the Lean theorem prover. Capability demonstrated in a paper before a model existed to sell was an unusual way to introduce a flagship, and it set the terms for the four weeks that followed.

The launch table led on reasoning: 99.9% on ARC-AGI-3, against the 30.2% Claude Opus 5 had posted on the same interactive-environment board in July, 97.6% on the rebuilt second edition of FrontierMath Tier 4, and 96.0% on GPQA Diamond. Agentic work moved by similar margins — 64.6% on Terminal-Bench Science 0.1 and 41.4% on AutomationBench, both more than double GPT-5.6 Sol's figures in the same run, and 74.1% on DeepSWE v1.1. The security rows are the ones that explain the wait: 100.0% on ExploitBench, the Carnegie Mellon capability ladder that scores how far a model gets turning a known V8 bug into a working exploit, and 99.2% on SRE-Bench over four attempts.

The professional and coding tables ran closer. At launch, OpenAI reported Astra at 57.9% on Terminal-Bench 4.0 against Claude Fable 5.1's 55.8%, BrowseComp at 91.5%, and 0.84 on the OpenScore String Quartets sheet-music transcription set where GPT-5.6 Sol had managed 0.19 — but on FrontierCode 1.1 its 53.3% on the main split and 64.5% on the extended set both sat fractionally behind the Claude models OpenAI ran beside it, and it trailed on Artificial Analysis' two composite indices: 61.2 on the Intelligence Index against Fable 5.1's 65.7, and 67.0 on the Coding Agent Index. For a generational release the shape of the table was lopsided — enormous margins on reasoning, security and the internal professional sets, and a coding field the previous generation had already crowded.

That capability is what OpenAI spent August evaluating. On August 7 the company said it had paused parts of the work on Astra because preliminary results could not rule out that the model reached the Critical cybersecurity level in its Preparedness Framework — the threshold describing a model able to find and carry out attacks against well-defended systems without human help, and the first time OpenAI had put one of its own models in that bracket. It returned to the question on September 1 in "Path to Astra: critical capabilities and frontier safeguards", saying the further evaluations were complete and the safeguards it had built around the model were sufficient to ship. Astra launched two days later.

Frequently asked questions

GPT-6 Astra was released by OpenAI on Thursday, Sep 3 2026.

All OpenAI releases

47 tracked

2022

1 release
GPT-3.5
Nov 30 2022