GPT-6 Astra Lands: A Model That Works Your Computer

OpenAI's newest model is not built to write about your work. It is built to sit at your computer and do it.

What happened

OpenAI released GPT-6 Astra on 03/09/2026. Access began the same day with a limited set of organisations, then widened from 04/09/2026 to the OpenAI API and to paid ChatGPT plans covering Plus, Pro, Business and Enterprise. Microsoft Foundry listed it as generally available on 03/09/2026 and GitHub Copilot followed on 04/09/2026. One detail worth noting for larger firms: on Enterprise plans, access is switched off by default, so an administrator has to turn it on deliberately.

The claim OpenAI leads with is computer use. On the OSWorld 2.0 benchmark it reports 72.6 per cent for Astra against 65.7 per cent for GPT-5.6 Sol, and says Astra reaches that score in roughly 47 per cent less time per task. On ScreenSpot-Pro, which tests whether a model can find the right control on a screen, it reports 92.7 per cent against 76.9 per cent. Translated out of benchmark language, this is a model designed to drive software rather than describe it: filling in forms, updating customer records, pulling figures out of a dashboard, formatting a document to an existing template.

Price is the other headline. The API lists at 10 US dollars per million input tokens and 50 US dollars per million output tokens, which is two and a half times the rate for GPT-5.6 Sol. Requests above 272,000 input tokens are billed at double the input rate and 1.5 times the output rate across the whole request. Astra usage sits inside existing ChatGPT subscription allowances, with extra credits available to buy.

Why it matters for UK businesses

Most small and mid sized firms in Britain do not have a clean API for the systems that actually eat their week. The supplier portal has no integration. The accounting package exports a spreadsheet and nothing else. The council planning site is a form somebody fills in by hand. Those gaps are exactly what a competent computer-use model closes, because it works the same way a person does, through the screen.

That changes the shape of automation projects. Until now, the honest answer to "can we automate this?" often depended on whether the vendor offered a connector. If a model can operate the interface reliably, the connector stops being the gate.

There is a second point UK firms should read carefully. OpenAI states that Astra is its first model to reach the Critical level for cybersecurity capability under its own preparedness framework, and it scored 100 per cent on ExploitBench compared with 78.5 per cent for the previous model. OpenAI has restricted the launch version so it refuses advanced offensive tasks, and has added misalignment monitoring that can pause or stop a job mid-run. That is the right call, but it means a task can be halted by a safety check rather than by a bug, and your process needs to cope with that.

How Adevious AI sees it

We would treat Astra as a challenger, not a replacement. The place we would put it first is the narrow, repetitive, screen-bound job that nobody wants: rekeying supplier invoices, reconciling two systems that refuse to talk, chasing a status across three portals. Those tasks have a clear right answer, so you can check the work.

What we would not do is hand it an unsupervised path to anything that spends money, sends a message on your behalf, or changes a record you cannot easily restore. A model that operates your screen inherits whatever permissions the logged-in account holds, which makes account hygiene a real control rather than a tickbox.

The caution to weigh is cost behaviour. At two and a half times the previous rate, and with a long-context surcharge on top, a task that loops or retries can turn expensive quietly. OpenAI says Astra uses fewer output tokens per task, which may offset the per-token price, but whether it does depends entirely on your workload. Benchmarks published by a vendor are evidence, not a forecast for your business.

One thing to do this week

Pick one screen-bound task, time it honestly, and write down what it costs you per month in labour. Then run the same task through Astra on a paid plan for a fortnight, logging every failure and every intervention. You will end up with a real number rather than a hunch, and that number is what should decide whether this gets a budget line.

If you want a second opinion on which of your processes are worth testing first, or you would rather someone else ran the fortnight for you, we are happy to have that conversation. Get in touch with Adevious AI and we will look at it with you.

Sources: https://openai.com/index/gpt-6-astra/ https://openai.com/index/safety-overview-gpt-6-astra/ https://developers.openai.com/api/docs/models/gpt-6-astra https://azure.microsoft.com/en-us/blog/gpt-6-astra-frontier-intelligence-for-work-now-generally-available-in-microsoft-foundry/ https://github.blog/changelog/2026-09-04-gpt-6-astra-is-generally-available-in-github-copilot/

Acronyms: API: Application Programming Interface UK: United Kingdom