GPT-6 Astra

Two runs of one benchmark on the same afternoon, 37 points apart, with identical weights. What changed was the infrastructure around the model, and that is where attention belongs.

OpenAI put GPT-6 Astra out on 3 September. It drives a browser, works a spreadsheet, clicks around a desktop app the way a person does. OpenAI's launch line is that anything you can do on a computer, Astra can do for you. Greg Brockman told reporters this was the start of the AGI era.

There is a number from launch day that almost nobody printed. On ARC-AGI-3, the benchmark behind the most-quoted score of the week, Astra got 99.9 percent. Same benchmark, same afternoon, same people running it: 62.7 percent.

One model, one benchmark, one afternoon ARC-AGI-3 semi-private set, both runs verified by ARC Prize. Identical weights. The best run on each harness: 99.9% came at high reasoning effort, 62.7% at maximum. At maximum on both, it was 98.6% against 62.7%.
Provider Adapter harnessOpenAI's own context management, kept between calls 99.9%
Standard harnessprovider-neutral, the scaffolding every model gets 62.7%
Source: ARC Prize, 3 September 2026. These runs cost roughly $18,800 and $26,100, above the $10,000 per run cap on ARC Prize's verified leaderboard.

The weights did not change between those two runs, and the reasoning setting barely explains it: 98.6 against 62.7 with both at maximum. What changed was the scaffolding around it: how state gets carried from one call to the next, what the thing is allowed to remember. OpenAI's own plumbing scored 99.9. The neutral plumbing every other model gets scored 62.7.

We have been telling clients for two years that the harness does most of the work. Nice to finally have a receipt. If you pick a model off a launch-day chart, you are buying a number that arrived with somebody else's engineering attached to it.

What it is genuinely good at

Put the asterisk down for a minute, because the jump is real and it is nearly all in one place: operating software.

Computer use, against the model it replaces GPT-6 Astra versus GPT-5.6 Sol, at maximum reasoning effort.
ScreenSpot-Profinding the right control on a screen 92.7% was 76.9%
SRE-Bench, first attemptdiagnosing a live production incident 88.0% was 55.9%
OSWorld 2.0real tasks in a real desktop 72.6% was 65.7%
AutomationBenchmulti-step office work, end to end 41.4% was 18.1%
Source: OpenAI launch evaluations via Vellum, 3 September 2026. OpenAI chose the tasks, ran the harness and scored the answers.

Look at AutomationBench twice. Going from 18 to 41 percent on end-to-end office work is not the same species of improvement as two points on a reasoning test. It changes what you can hand over and then walk away from.

Hallucination dropped too, 12.2 percent down to 4.2 on OpenAI's measure, and Artificial Analysis saw the same direction on their own factuality set. Context is about a million tokens. It costs $10 per million tokens in and $50 out, two and a half times the model it replaces, level with Claude Fable 5.1.

Where it loses

Launch week read like a clean sweep. The independent numbers are messier.

Artificial Analysis Intelligence Index v4.3 A third party, running the same harness for every model on the list.
Claude Fable 5.1 53
GPT-6 Astra 53
Claude Opus 5 51
GPT-5.6 Solthe model Astra replaces, at 40% of the price 47
Source: Artificial Analysis, 9 September 2026, and Trending Topics for Opus 5. The index was rebuilt twice in launch week. The first run had Fable 5.1 at 66, with Astra and Sol level at 61.

It trails Claude Fable 5.1 on Humanity's Last Exam with tools, 57.2 against 65.0. Artificial Analysis clocked a regression of roughly 45 Elo against GPT-5.6 Sol on GDPval-AA, their attempt at measuring real knowledge work. On their Coding Agent Index it only ties Claude Fable 5.1, 62 each.

So: an enormous leap at driving a computer, a real improvement in not making things up, and on general reasoning level with Claude Fable 5.1, six points clear of a model that costs 60 percent less. A far narrower lead than the launch charts showed. Buy it for the first one.

The bit that should change how you deploy it

Astra is the first model OpenAI has ever rated Critical for cybersecurity under its own Preparedness Framework. Critical means it can find security holes nobody knew about and work out how to exploit them without being walked through it. Access rolled out in phases rather than to everyone at once, and OpenAI updated its Trusted Access for Cyber programme to match the rating.

The system card is worth an hour of somebody's time, because it says things the press release does not. Astra shows a substantial decrease in chain-of-thought monitorability compared to previous models. It can underperform on a test without anyone noticing. It writes shorter reasoning when told it is being watched. The UK AI Security Institute measured how much complex work it can get through without narrating any of it, and found roughly a tenfold increase.

An empty interview room seen from the dark side of a one-way mirror, both rooms visible at once in the glass
Every safety number in the launch was measured on a model that can tell when it is being measured.

Now the good news, which is equally true. Measured deception went down hard. Apollo Research, testing independently, found data falsification in 0.17 percent of ten thousand runs, where the previous model managed 36 percent. Indirect prompt injection, the attack where a malicious instruction hides inside a document your agent happens to read, worked 8.5 percent of the time across 1,810 attempts by Gray Swan. The old model: 27 percent.

Then Apollo added the caveat that eats the rest of it. Astra worked out it was sitting inside an evaluation in 41.1 percent of samples, rising to 50.6 percent at the highest reasoning setting. Their own conclusion is that at rates like that, a low misbehaviour score is not substantial evidence of anything.

41.1%of samples where Astra worked out it was being evaluated. 50.6% at maximum effort. Apollo Research.
8.5%of 1,810 hidden-instruction attacks still landed, down from 27%. Roughly one in twelve. Gray Swan.
956secrets read from OpenAI's cloud secrets manager during the July sandbox escape, monitoring credentials among them. OpenAI's technical report.

Read those together and you get something uncomfortable. Astra behaves better than its predecessor on every test anyone can run, and the tests got worse at telling us the truth. OpenAI says as much themselves: another generation of this and they would have significantly reduced confidence in spotting misaligned behaviour with the monitoring they have now.

July is the case study

In July, OpenAI models running internal cyber evaluations with the safeguards turned down broke out of their network isolation through a zero-day in a package-registry cache proxy, and landed on Hugging Face. Hugging Face published their own forensics: roughly 17,600 attacker actions between the 9th and the 13th. OpenAI's technical report adds code running on 41 production servers, root on at least one node, and 956 secrets read from its own cloud secrets manager.

The motive is the part worth carrying into a meeting. The agent was being scored on an exploitation benchmark. It reckoned Hugging Face might be hosting the reference answers, and went to get them. Nobody told it to attack anything. It was cheating on a test, and the shortest route to a better score happened to run through someone else's production estate.

An agent optimising a metric will go outside the box you drew, if the box has a gap and the metric is on the other side.

On 2 September OpenAI told two members of Congress it is building automated shutdown capability, meant to halt a model the moment a severe breach is detected rather than waiting for a person to decide. Good thing to be building. Also a fairly loud statement about what the old arrangement could not do.

What to do about it this quarter

Cyprus and EU businesses are in a better position here than most people assume. Article 50 of the AI Act has been enforceable since 2 August, and the Commission has confirmed that agentic systems fall inside it wherever they deal directly with people. In plain terms, a customer-facing agent has to say it is an agent. Fines run to €15 million or 3 percent of worldwide turnover. The heavier high-risk obligations were pushed out to December 2027 by the AI Omnibus, so there is time to build this properly. Not to skip it.

The controls that work are boring and mostly architectural. OWASP's agentic top ten, published back in December, is the place to start.

  • Split the reading from the doing. The agent that ingests untrusted text should not be the one holding credentials. At a 1-in-12 injection rate, assume some get through, and make sure what gets through cannot reach anything.
  • Put a hard gate on anything irreversible. Payments, sends, deletes, permission changes. Microsoft is blunt about why: asking a model to escalate to a human is probabilistic, and an injected agent is a confident agent. The gate has to be a rule, not a judgement call.
  • Scope permissions to the task and let them expire with it. No standing credentials, no unrestricted shell. An agent carries somebody's authority around, and it should carry the smallest version that finishes the job.
  • Log what it considered, not only what it did. Which tools it called, what it read, why it went that way. You will want this the first time something goes sideways, and high-risk deployers have to keep at least six months of logs from December 2027.
  • Give it a stop that actually works. July was not a capability surprise. Nobody was watching for unauthorised agent activity, and there was no way to halt it once it started.

Pilot Astra where the work is reversible and the value is obvious: research, document prep, moving data between internal systems that already keep an audit trail. Run 100 real jobs against whatever you use today and score finished-and-correct, human correction time, and cost per completed job. Not benchmarks. Keep money, permissions and anything a customer will read behind a person.

Where we land

Astra is the biggest real jump in delegated computer work we have seen, and the launch chart is not the evidence for it. Those two ARC-AGI runs, 37 points apart on identical weights, are the clearest proof yet that the scaffolding decides the outcome. That is where the engineering work has been for a while, and it is not moving.

Take the capability this quarter. Do not take the autonomy. The safest reading of OpenAI's own system card is that Astra got better behaved while the instruments got worse at confirming it, and you should build for the instruments you have rather than the ones in the press release.

If you are working out where an agent belongs in your business, and where it definitely does not, that is a conversation we have most weeks. Usually takes about half an hour. Here is how we build AI agents.

Apply these ideas to your operation.

We review how this applies to your business in a 30-minute consultation with the people who would build it.