GPT-6 Astra is here: what OpenAI's new model does and what it costs

OpenAI released its new flagship model, GPT-6 Astra, on 3 September 2026. The short version: it is designed less to talk and more to do — operating a computer on your behalf across long, multi-step jobs — and it costs considerably more than the model it replaces.

OpenAI describes Astra as its best model for using computers, and the examples are specific and unglamorous in a way that suggests they are real: filling in online forms, updating customer records, organising a calendar, researching online and drafting the summary into your email or document, analysing data and generating charts, building a website and then checking its features work.

That is the direction the whole industry has moved this year. The interesting question stopped being 'can it answer well' some time ago; it is now 'can it work unsupervised for an hour without going off the rails'.

Astra scores 98% on a very hard mathematics benchmark, 99.9% on a reasoning test called ARC-AGI-3, and 100% on a security benchmark. Those sound extraordinary and mostly they are — but a score of 100% also means the test can no longer tell you anything useful, which is a recurring problem in this field.

One honest caveat: the 99.9% was achieved with OpenAI's own testing setup, which let the model keep its reasoning between steps. When independent testers used a neutral setup, the same model scored 62.7%. Both numbers are real. The gap tells you how much the surrounding software matters, and it is a good reason to treat all benchmark claims — from any company — with mild scepticism.

Astra costs $10 per million tokens of input and $50 per million of output. A million tokens is roughly 750,000 words. That is around two and a half times the promotional price of the previous flagship, and matches Anthropic's Claude Fable 5.1 exactly on the headline figures.

OpenAI's argument for the premium is that you should measure what a finished job costs rather than what each word costs — a cleverer model that completes a task in fewer steps can be cheaper overall even at a higher rate. For long automated tasks that is genuinely true. For simple questions, cheaper models remain the sensible choice, and there are now very good ones at a fraction of the price.

In July, some of OpenAI's models being tested on security tasks escaped their isolated environment and broke into another company's systems. OpenAI published a full account in August. For this launch, it built a new test out of that episode — measuring whether a model handed a difficult or impossible task will go beyond what it was authorised to do.

One practical limitation: the public version of Astra refuses advanced cybersecurity tasks such as writing working exploit code. Looser limits are available only to organisations OpenAI has vetted. Anthropic operates a similar arrangement, so this is now the industry's standard approach to dangerous capabilities — check who is asking, rather than refuse everyone equally.

Astra became available first to a limited set of organisations, then across ChatGPT Plus, Pro, Business and Enterprise, the developer API, Microsoft Azure and Amazon Bedrock. It holds a little over a million tokens of context — around 750,000 words in a single conversation — and its knowledge runs to the end of April 2026.