GPT-6 Astra: What the Benchmarks Actually Show
OpenAI released GPT-6 Astra on September 3, 2026 as a limited preview, with a wider rollout the following day. (If you've been hearing it as "GPT Astro," the name is Astra.) The announcement language was maximal: the world's most intelligent and aligned model, a generational leap, state of the art on computer use, browsing, software engineering, cybersecurity, science, and professional work.
Some of that is true in a way that matters. Some of it is a headline number doing a lot of work. The interesting thing about this launch is that the two categories are easy to separate if you look at the actual scoreboard — so let's do that.
Where Astra genuinely jumped
The gains are real and concentrated in three places: math, agentic computer use, and offensive security.
| Benchmark | Astra | Prior / rival |
|---|---|---|
| FrontierMath Tier 4 | 97.6% | 87.8% (Fable 5.1) |
| ARC-AGI-3 | 99.9% | 30.2% (Opus 5) |
| GPQA Diamond | 96.0% | highest published |
| OSWorld 2.0 (computer use) | 72.6% | 65.7% (GPT-5.6 Sol) |
| ScreenSpot-Pro | 92.7% | 76.9% (Sol) |
| ExploitBench | 100% | 78.5% (Sol) |
The computer-use numbers are the ones I'd actually plan around. ScreenSpot-Pro going from 76.9% to 92.7% is the difference between an agent that misclicks often enough to need supervision and one that mostly doesn't. OpenAI also says Astra is roughly twice as fast at computer use, finishing OSWorld tasks about 47% quicker. For anyone building agents that drive real software, latency and click accuracy are the product — that combination is a bigger deal than any single reasoning score.
The security result deserves its own sentence. Astra scored 100% on ExploitBench, and on an internal V8 vulnerability port it achieved arbitrary code execution 39.0% of the time versus Sol's 5.5%. It also surfaced two genuine zero-days during evaluation. That is why the advanced cybersecurity capability ships gated to a limited set of vetted testers, and why enterprise access is off by default until an admin turns it on.
Where the story gets quieter
Now the part the launch post skipped. On coding — the thing most readers of this blog actually care about — Astra barely moved.
| Benchmark | Astra | Field |
|---|---|---|
| DeepSWE v1.1 | 74.1% | 72.7% Sol · 73.7% Opus 5 · 73.8% Gemini 3 Flash · 75.4% Muse Spark 1.3 |
| FrontierCode Extended | 64.5% | 64.9% Fable 5 |
| Humanity's Last Exam | 57.2% | 65.0% Fable 5.1 |
| Intelligence Index (independent) | 61 | 61 GPT-5.6 Sol — no change |
DeepSWE is the coding benchmark that tracks real developer experience most closely, and Astra lands 1.4 points above its own predecessor while sitting below Meta's Muse Spark 1.3. On FrontierCode Extended it trails Fable 5. On Humanity's Last Exam it doesn't just fail to lead — it regresses relative to a competitor by almost eight points. And the independent Intelligence Index puts the aggregate at 61, identical to GPT-5.6 Sol.
A "generational leap" that leaves the composite score unchanged is a specific kind of release: a model that got dramatically better at a few things and stayed put on everything else.
Three caveats worth carrying
The scores were run at maximum effort. Unless noted otherwise, the published numbers come from max-effort configurations. That lifts accuracy while inflating latency and token consumption — so the benchmark model and the model you get at default settings are not the same thing, and your bill won't look like the benchmark either.
Some of the biggest wins are on home turf. OpenAI funded FrontierMath's development, and it runs its own harness for ARC-AGI-3. That doesn't make 97.6% or 99.9% fake, but a vendor-influenced benchmark is weaker evidence than an independent one, and those two happen to be the headline figures.
The new architecture is harder to watch. Astra uses a "recurrent depth" or looped-transformer approach to reasoning, which is part of why it's more efficient. It's also why safety researchers have flagged monitorability concerns: the technique obscures some or all of the model's reasoning trace. OpenAI reports a better honeypot result (0% versus 48.2%), but "more aligned" and "harder to inspect" are being claimed at the same time, and those pull in opposite directions.
The Codex change you might actually feel
Buried under the benchmark noise is a practical one. Astra changes how long sessions carry state: rather than compressing prior work into a lossy summary when the context window fills, it keeps notes across windows, preserving accumulated detail while leaving earlier conversation searchable.
If you've ever watched a long agent session quietly forget a constraint you set an hour ago, you know why this matters more than a point of DeepSWE. It's the difference between an assistant that degrades over a workday and one that doesn't.
What it costs
API pricing is $10 per million input tokens and $50 per million output — the same list price as Anthropic's Claude Fable 5.1 as of launch day, and about 2.5x what Sol was going for on promotional rates. Fast mode runs up to twice as quick at twice the price. It's included in Plus, Pro, Business, and Enterprise subscriptions with extra credits purchasable, and it's available through the OpenAI API, Microsoft Azure, and AWS Bedrock.
So should you switch?
It depends entirely on what you're doing, which is the actual takeaway here.
- Building agents that operate software? Yes — this is the clearest upgrade in the release, and the speed gain compounds with the accuracy gain.
- Hard math or research-grade reasoning? Yes, with the caveat about who funded the benchmark.
- Day-to-day coding? Probably not worth a migration on these numbers. The field is bunched within about two points on DeepSWE, and Astra isn't at the top of it.
- Long agent sessions? Worth testing for the context-notes change alone, independent of any benchmark.
The honest summary of GPT-6 Astra is that it's a strong agentic release wearing the marketing of a general intelligence leap. Those are different products, and the second framing is the one that will age badly. Worth being precise about which one you're buying.