GPT-6 Astra: The Future Is Here
GPT-6 Astra is OpenAI's most intelligent and aligned model yet — a frontier system welcomed by many as the opening of the AGI era. It matches human experts on frontier benchmarks, operates computers like people do, and writes production-grade software.
The AGI Era
Welcome to the AGI era
Within a day of launch, GPT-6 Astra moved the AGI conversation from theory to front pages. On ARC-AGI-3 it used fewer actions than the median human tester on 96% of levels — what ARC Prize calls "effectively reaching human parity on the benchmark". Here is what independent evaluators are saying.
“The best model we've ever tested.”
“The story is: end of one era, start of another.”
“A noticeable step-function change in frontier model capabilities.”
Is it actually AGI? The honest answer.
- OpenAI markets Astra as "the world's most intelligent and aligned model" — a new generation of intelligence, and its closest step toward AGI.
- ARC Prize, whose benchmark Astra essentially saturated, is explicit: "we are not claiming that it is AGI" — deterministic benchmarks don't capture real-world open-endedness.
- What's not in dispute: the frontier moved further in one release than in years — and the future it points at is no longer hypothetical.
Capabilities
One model, every frontier
Astra pairs state-of-the-art raw intelligence with the practical ability to do professional work end to end.
Computer use
Astra sees the screen and works like a person: filling forms, updating CRMs, running frontend QA and installing software — finishing OSWorld tasks in about 47% less time than GPT-5.6 Sol.
Best-in-class coding
Described by OpenAI as the best software engineering model to date. A new Codex integration keeps searchable notes across context windows instead of compressing them away.
Science & math
Astra helped establish a new bound of 186 for short prime gaps — improving a term in the mathematics that had stood unchanged for over 80 years.
Million-token memory
Reads entire codebases and book-length corpora in one pass, scoring 96.3% on MRCR v2 (8-needle) across 512K–1M-token evaluations.
Agentic workflows
Plans and executes long professional tasks end to end. Partners report complex creative workflows completing with up to 20% fewer tokens than other frontier models.
Alignment & safety
On an impossible-task eval, Astra violated its scope 0% of the time vs 48% for GPT-5.6 Sol. During testing it discovered two real zero-days — and disclosed them to maintainers.
Benchmarks
The numbers behind the moment
GPT-6 Astra vs. its predecessor GPT-5.6 "Sol" and the strongest competing models at launch.
| Benchmark | GPT-6 Astra | GPT-5.6 Sol | Best other |
|---|---|---|---|
| ARC-AGI-3 (adapter harness) | 99.9% | 7.8% | 30.2%Claude Opus 5 |
| FrontierMath Tier 4 | 97.6% | 83.0% | — |
| ExploitBench | 100% | 78.5% | — |
| SRE-Bench (single try) | 88.0% | 55.9% | — |
| Terminal-Bench 4.0 | 57.9% | — | 55.8%Fable 5.1 |
| GPQA Diamond | 96.0% | 94.6% | — |
| Agents' Last Exam | 59.3% | — | 55.5%Claude Opus 5 |
ARC-AGI-3's 99.9% is under the provider-adapter harness (62.7% under the standard harness — both state of the art). On action efficiency, Astra used ~51.7% fewer actions per level than the median human tester. Scores as reported by OpenAI and ARC Prize, September 2026.
Specifications
Under the hood
Key facts for builders evaluating the API and teams choosing a plan.
FAQ
Frequently asked questions
Deep dives
Read the primary sources
Everything on this page is drawn from official releases and independent evaluations.
The future is here. Go meet it.
GPT-6 Astra is rolling out now across ChatGPT, the OpenAI API, Azure and AWS Bedrock.
This page is an independent overview published by Findry AI, an AI tools directory. GPT-6 Astra is a product of OpenAI — benchmark figures as reported in the sources above, September 2026.