GPT-6 Astra: OpenAI's new flagship model

OpenAI unveiled GPT-6 Astra on 3 September 2026. The promise isn’t that it writes better prose. It’s that the model operates software the way a person would: working in the browser, in spreadsheets and in desktop applications, producing finished documents and presentations, and seeing multi-step tasks through on its own. OpenAI calls it the best model available for exactly this kind of computer use.

It launched first to a limited number of organisations; over the coming days it should reach the ChatGPT Plus, Pro, Business and Enterprise plans. It is also available through the OpenAI API, Microsoft Foundry and AWS. Here’s what that changes – and what it doesn’t.

 

InfoBenefit: Perspective, not headlines
UseCaseUseCase: Planning your model strategy
ZeitReading time:
4 min
SchwierigkeitDifficulty: Easy

The useful question isn’t how good Astra is, but what it’s good at. Its strength is operating software by itself and seeing multi-step tasks through: the kind of work where, until now, someone had to click through it, check it and hand it on. For writing, summarising and the everyday office work most people do, not much changes. Your current model is still the sensible choice for that.

You’ll find Astra in ChatGPT, in the OpenAI API and in whatever you build on Microsoft Foundry. You won’t find it in Microsoft 365 Copilot, which has been running GPT-5.6 since July with no successor announced by either company – Microsoft’s post about Astra arriving in Foundry doesn’t mention Copilot once. If you’re planning work there, carry on with GPT-5.6.

Then there’s the headline that was everywhere this week: Astra is the first OpenAI model to reach “Critical”, the top level for cybersecurity capability in the company’s own Preparedness Framework. In practice that means it can find unknown security flaws by itself, given the right tools, and work out how to exploit them; in testing it found two previously unknown vulnerabilities and escaped the test environment. For everyday work, though, the context matters more than the headline: the model as shipped refuses offensive tasks and supports defensive work only, and OpenAI says those record cyber scores come from a setup without those safeguards. The rating is also a self-assessment, measured against a framework the company wrote itself.

  1. ChatGPT Plus, Pro, Business and Enterprise: Astra is rolling out to these plans “over the coming days” – OpenAI gives no firmer date than that. The model picker will tell you whether it has reached you.
  2. Off by default in Business and Enterprise workspaces: press reports say Astra is disabled at launch and that an administrator has to enable it for the workspace.
  3. Through the OpenAI API and AWS: 10 US dollars per million input tokens and 50 per million output tokens; a “fast mode” costs double for up to twice the speed.
  4. Through Microsoft Foundry: initially via the Limited Access Program, billed by consumption, with no reserved capacity required. Prices run from 10 to 22 US dollars per million input tokens and 50 to 82.50 per million output tokens, depending on context length and region.
  5. Extended cyber capabilities in stages only: these go first to a small group of alpha testers, then through the Daybreak Blue programme, for defensive uses such as vulnerability checking and malware analysis.

Independent hands-on tests are still to come – the model is too new for that. So far there are two customer stories, both published by OpenAI, and they should be read in that light. The legal software firm Legora had Astra review 41 documents in a matter of minutes; it found all four errors deliberately planted in the accounts, among them a £500,000 gap in the revenue note, and ran around 50 more checks than the previous model managed. The games developer Playco turned a single unfinished prototype into three fully themed variants in one pass, and reports roughly 50 per cent less manual rework.

One early independent measurement comes from the analysis firm Artificial Analysis. On its broad comparison index Astra scores 61.2, almost level with GPT-5.6 Sol on 60.9 – the clearer lead shows up in agentic coding tasks, at 67.0 against 65.1. It also costs noticeably more to run. That puts the leap the announcement implies into perspective, without diminishing the real progress on software development.

  • The benchmark figures come from the vendor: one example shows how much the method matters. On ARC-AGI-3, OpenAI reports 99.9 per cent using its own test setup – on the ARC Prize organisation’s standard setup it is 62.7 per cent.
  • The monitoring can interrupt your work: OpenAI checks every tool call automatically for unwanted behaviour, and says itself that the system may wrongly flag legitimate activity, which can leave real work “slowed, paused, or stopped”. For production use that is a real operational factor.
  • No EU data zone named: for Foundry, Microsoft lists deployments in “Global” and in the US data zone. It doesn’t name an EU data zone – something to settle in advance for projects with those requirements.
  • Work out who actually needs it before switching it on: Astra’s strength is operating software by itself. For everyday writing and summarising it adds little, and your existing model still does that job.
  • Keep an eye on cost: Foundry bills by consumption, and the prices sit well above the previous generation. Set a fixed budget for early experiments and pick one clearly bounded use case.

It isn’t new that models are judged on more than how well they write – this release just makes it unusually plain. The question now is whether a model carries a task through to the end: clicking, checking, correcting, delivering. There are benchmarks built for exactly that now, and they are designed to mimic real workflows. But unlike writing, where the test and the real thing are essentially the same act, an agentic workflow inside your own organisation can fail in places no test setup knows about: legacy systems, permissions, edge cases. That makes the numbers a good indication – but an indication.

And when numbers are only indications, who interprets them matters all the more. OpenAI president Greg Brockman told a press briefing it was “not unreasonable to feel that we are now in the AGI era” – the words of a company representative, not a finding. Most of the evidence behind that claim comes from OpenAI itself; the first independent measurement puts Astra roughly level with its predecessor on average, and clearly ahead only on coding tasks. What the model does in real workflows will become clear over the next few weeks. Until then, the staged access and the pricing are reason enough to start small.