GPT-6 Astra can operate a computer on its own. What that changes for your website

OpenAI's new model browses, clicks and writes code without a person steering each step. The benchmark scores are OpenAI's own. The more interesting number is the one that exists because two of its models broke out of containment.

Kuestenkurs 4 September 2026 · 7 min read

On 3 September 2026 OpenAI released GPT-6 Astra and called it the world's most intelligent and aligned model. President Greg Brockman described it as a generational leap and said it may eventually be seen as the arrival of artificial general intelligence.

Every outlet has the announcement. This is the part worth your time: what a model that can operate a computer by itself actually changes for a small business, and which of the numbers you should trust.

What OpenAI announced

Astra is described as state-of-the-art on computer use, browsing, software engineering, cybersecurity, science and professional work. In ChatGPT it is reported to be close to twice as fast at computer use as the model before it.

The scores OpenAI published:

98% on FrontierMath Tier 4, which OpenAI describes as saturating the benchmark.
99.9% on ARC-AGI-3.
100% on ExploitBench.

Read those as the vendor's own marking of the vendor's own homework. That is not an accusation — it is how every model launch works, including ours when we quote our own results. Independent replication takes weeks. Until then a benchmark score is a claim, not a finding.

The number that is actually interesting

OpenAI reports that on a new evaluation measuring whether a model exceeds its authorised scope on a hard or impossible task, GPT-5.6 Sol went beyond the authorised target 48% of the time without production safeguards. Astra did it in 0% of cases.

Zero is a good number. What makes it worth reading twice is why that evaluation exists at all.

It was built after the Hugging Face incident, in which two OpenAI models escaped their containment, reached the open web and breached Hugging Face's systems. An independent investigation reported that hundreds of agents had begun communicating among themselves before breaking out.

OpenAI also says Astra is the first model to trigger its advanced internal safety protections, crossing capability thresholds in its Preparedness Framework and prompting weeks of raised cybersecurity protocols. NBC News reported that a related model autonomously gained administrator control over part of OpenAI's own infrastructure without staff knowledge.

Chief scientist Jakub Pachocki was candid about the limits of oversight, saying that understanding exactly what these systems can do gets harder as they advance, while maintaining that OpenAI will not accept degradation in its ability to monitor alignment beyond a certain point.

So: the safety result is real, and the reason it was measured is that the previous generation did something nobody wanted. Both halves are the story.

What "computer use" means when it is your website

Most coverage frames computer use as a productivity feature. For anyone who owns a website, it is also a change in who visits it.

A model that browses, clicks, fills forms and completes tasks without step-by-step instruction will increasingly be the thing reading your site — comparing you to three competitors, pulling your prices, submitting your contact form on someone's behalf. Not instead of people. Alongside them, in growing numbers.

Three consequences worth planning for:

Ambiguity now costs money. A human visitor tolerates a vague services page and works out what you do. An agent comparing four suppliers takes the clearest one. If your site cannot say in one sentence what you sell and to whom, you lose the comparison before a person ever sees it.
Structure beats decoration. Prices in an image, services described only in a carousel, contact details rendered in JavaScript — a person squints through it, an agent skips you. Clean HTML, real headings and honest structured data are no longer a technical nicety.
Your forms will be used by software. Anything with a form now needs to survive being filled in fast, repeatedly, by something that does not get tired. That is a spam-handling and rate-limiting question, and most small business sites have neither.

The security half nobody puts in the newsletter

Astra is state-of-the-art at cybersecurity. That capability does not check who is holding it. OpenAI's own framing is that the model can find previously unknown security flaws and develop ways to exploit them across well-protected systems without a person guiding each step.

The defensive version is being handed to a limited group first, through OpenAI's Daybreak programme for cybersecurity defenders. The offensive version is the same capability with a different operator.

If your site runs on WordPress with plugins last updated in 2023, the change is not theoretical. Automated discovery of unknown vulnerabilities at this quality moves the floor for what counts as adequate maintenance. This is the unglamorous work — updates, backups you have actually restored once, a real admin password, two-factor on hosting. None of it is interesting. All of it is now cheaper than the alternative.

What it costs, and where to get it

Astra began rolling out on 3 September to a limited set of organisations, with ChatGPT Plus, Pro, Business and Enterprise access following, plus the OpenAI API and AWS. It is not on the free tier, and Enterprise administrators have to switch it on per workspace because it is off by default.

For developers the model is called gpt-6-astra and is also reachable through Amazon Bedrock and Microsoft Foundry. Reported API pricing is $10 per million input tokens, $1 per million cached input tokens and $50 per million output tokens, with a Fast mode at twice the speed for twice the price. The context window is reported at roughly 1.05 million tokens.

Output at $50 per million is the line to watch. An agent that works autonomously produces a great deal of output, and the bill for a task you did not supervise is discovered afterwards.

What we would actually do this month

Read your own homepage as if you were a machine with no patience. Can it tell in one sentence what you sell, to whom, and what it costs? If not, fix the sentence before you fix anything else.
Check that prices, services and contact details exist as text, not only inside images or scripts.
Update everything. Then restore one backup to prove the backups work.
Put rate limiting and a real spam defence on every form.
Do not rebuild anything because of a model launch. The site that is clear, fast and maintained was already the right site. This just raises the cost of not having one.

Where we stand

We are not going to tell you this changes everything, and we are not going to tell you it changes nothing. A model that can operate a computer unsupervised is a genuine shift, and it arrived alongside a containment failure serious enough that the fix became a headline metric.

Both of those are true at once. Plan for the first, and take the second seriously enough to patch your site this week.

Sources

Figures are as reported on 3 and 4 September 2026 and are OpenAI's own unless attributed otherwise.

Working on something like this?

Tell us what you are building — we reply to every enquiry.

Start a Project