GPT-6 Astra: what it can actually do (with prompts to try)

OpenAI shipped GPT-6 Astra on 3 September 2026. In a briefing with reporters ahead of the release, president Greg Brockman opened with "Welcome to the AGI era" and said he personally believes OpenAI has reached AGI.^[5]

That's a big claim. So rather than relitigate it, here's a narrower question: what can this thing actually do that matters if you build software? Below is what the numbers support, organised by scenario, with a prompt to try in each — and a section on the three places the launch framing outruns the evidence.

The 60-second context

Astra replaces GPT-5.6 Sol as OpenAI's frontier flagship. Access reached ChatGPT Plus, Pro, Business and Enterprise over the days following launch, then the API, AWS, Azure and Amazon Bedrock.^[5]

The pitch is a model you hand whole tasks to rather than individual prompts — built for computer use, terminal workflows, cybersecurity and scientific research.

Scenario 1 — Computer and browser use

This is where Astra genuinely separates itself. On OSWorld 2.0, which measures how well an agent operates a real desktop, Astra scores 72.6% against Sol's 65.7%, with a 47% reduction in time per task.^[2] Higher accuracy and less wall-clock time is the difference between an agent you babysit and one you can walk away from.

On ScreenSpot-Pro, testing interface navigation, it hits 92.7% without external tools, against Sol's 76.9%.^[2]

"Open my browser, find the 3 cheapest flights from Kochi to Dubai next month, and put them in a table with dates and prices."

Scenario 2 — Terminal and agentic coding

Long, messy, multi-step terminal work is the other real strength. Terminal-Bench 4.0 covers software engineering, system configuration and data analysis in a terminal:^[3]

The lead over Fable 5.1 is under two points — inside the margin where you'd never notice in practice. The more interesting claim is efficiency: OpenAI reports Astra reaching that score with substantially fewer output tokens.

"Clone this repo, find why the test suite fails on Node 22, fix it, and re-run the tests until they pass."

Scenario 3 — Security review

Astra became the first model to meet OpenAI's own "Critical" cybersecurity threshold, meaning it can find previously unknown vulnerabilities and build working exploits without step-by-step human guidance.^[5] On ExploitBench it scored 100%, against Sol's 78.5%.^[3]

OpenAI's framing is defensive: a model good at finding vulnerabilities helps defenders patch first. Keep your own usage on that side of the line — auditing code you own, not probing systems you don't.

"Audit this authentication function for security vulnerabilities, rank them by severity, and show me the patched version."

Scenario 4 — Math and science

On FrontierMath Tier 4, Astra scores 97.6%, against 87.8% for Fable 5.1 and 73.2% for Opus 5.^[2] On Terminal-Bench Science, which tests whether an agent can run real research workflows, it reaches 64.6% against Fable 5.1's 52.6%.^[2]

One caveat worth knowing: Epoch AI, which builds FrontierMath, notes that OpenAI funded its development and has exclusive access to part of it.^[2] That doesn't make the score wrong, but it isn't an arm's-length exam either.

"Here's my experiment's raw CSV — fit the best model, run the stats, and tell me whether the result is significant. Show your work."

The reality check

Three places where the launch framing outruns the evidence.

The AGI score depends almost entirely on the harness. The 99.9% on ARC-AGI-3 requires OpenAI's stateful provider-adapter harness. On the standard stateless harness the same model scores roughly 17% to 63%, depending on reasoning tier.^[3] Same weights, wildly different story according to who runs the test — which is the single most important thing to understand about this launch.

Astra does not lead general coding. On DeepSWE v1.1 it posts 74.1% against Sol's 72.7% — but Meta's Muse Spark 1.3 scored 75.4% the same week, putting Astra second.^[2] The New Stack also points out that OpenAI's own comparison chart used a 67.4% Fable 5.1 result, which makes the lead look larger than broader results support.^[2] If your day job is shipping features and polishing front-ends, this is not a clear upgrade.

It's expensive — but not relative to everything. $10 per million input tokens and $50 per million output, about 2.5x Sol's $4/$20, with fast mode doubling again to $20/$100.^[2] Worth being precise about the comparison, though: that 2.5x is against its own predecessor. Against Claude Fable 5.1, which also charges $10/$50, it is a like-for-like rate — and Claude Opus 5 at $5/$25 is exactly half. I worked the per-task maths out in the full cost comparison. Brockman's counter is that "the price per task is what matters," and token efficiency does offset part of the gap, but launch-day data is too sparse to confirm it.^[2]

A fourth, less discussed: Astra is harder to audit. Its written reasoning is harder to monitor than Sol's on simpler tasks, because it solves problems in fewer written steps and has more control over what it records.^[5] For anyone who needs to explain how a decision was reached, that's a real trade-off rather than a footnote.

Bottom line

Reach for Astra when the job is computer use, terminal or agentic workflows, defensive security work, or long-horizon tasks where token efficiency pays for the premium.

Skip it when you're doing everyday coding, front-end work, or watching the budget. Opus 5 is within noise of the frontier on conventional coding at half the price, and Sonnet 5 or Haiku 4.5 handle high-volume work for a fifth or a tenth of it. Fable 5.1 matches Astra on capability and on price, so between those two the decision is capability, not cost.

The useful way to think about Astra isn't "the new best model." It's a specialist: clearly ahead on agentic and computer-use work, roughly level on conventional coding, priced at a premium, and carrying a headline reasoning score that mostly measures the harness. Match it to the job, not the headline.