OpenAI Has Released a Model That Uses Your Computer, and Priced It Like a Specialist
GPT-6 Astra arrived on 3 September with a million-token context window, a headline capability of operating a computer as a person would, and a price two and a half times its predecessor.
Outspoken Digest Technology Desk
Saturday, September 5, 2026/3 min read
OpenAI announced GPT-6 Astra on 3 September, describing it as the most intelligent and aligned model in the world. Public availability was set for 5 September, with the first access going to enterprise customers in the company's Daybreak cybersecurity programme.
The specifications are the least interesting part, and they are worth listing anyway because they set the frame for everything else.
What was announced
The API model identifier is gpt-6-astra. The context window is one million tokens. Pricing on the API is 10 dollars per million input tokens and 50 dollars per million output tokens, with cached input at 1 dollar, batch and flex processing at half rate, and a faster tier at double.
That is roughly two and a half times the price of the model it replaces. Rollout runs through the API, Amazon Web Services and the ChatGPT Plus, Pro, Business and Enterprise plans.
The capability OpenAI has put at the front is computer use: the model operating a machine the way a person would, through a screen and a pointer, rather than through a purpose-built interface.
Why computer use is the claim that matters
Because it changes what the thing is for.
A model that answers questions is a tool you consult. A model that can drive software is a tool you delegate to, and delegation is where the commercial argument for these systems has always lived. Almost all valuable office work runs through applications that were built for humans and have no useful programmatic interface. If a model can operate those applications directly, the addressable work expands enormously.
It is also the capability with the worst failure modes. A model that gives a wrong answer wastes your time. A model with a pointer and credentials can delete things, send things and buy things. The gap between those two risk profiles is the whole argument about agent safety, and it is why the company's framing of the model as aligned is doing so much work in the announcement.
The price tells you who it is for
Two and a half times is a substantial move in a market that has spent three years getting cheaper per token, and it says something specific.
This is not being positioned as the default model for everyone. At those rates, long agentic runs that consume large numbers of tokens become expensive quickly, and the economics only work where the task being automated is worth considerably more than the compute. Cybersecurity, software engineering and professional work, which is where OpenAI has aimed it, are exactly the categories where that is true.
The million-token context reinforces the reading. A context that large is for feeding a model an entire codebase, an entire case file or an entire corpus, and paying for the privilege.
The scrutiny
The launch has drawn immediate attention to safety, and the cybersecurity emphasis is the reason.
A model presented as state of the art at cybersecurity is, by construction, a model that is good at finding and exploiting weaknesses in software. Releasing it first to a controlled programme of enterprise security customers is a recognition of that, and it does not resolve it, because the same capability reaches a much broader audience within days.
This is the recurring shape of frontier releases now: the capability that makes the model commercially valuable and the capability that makes it dangerous are frequently the same capability, described in different words to different audiences.
What to actually watch
Not the benchmark table. Benchmarks at this level have become saturated and contested, and every lab reports the ones it wins.
Watch whether computer use survives contact with real software: legacy applications, inconsistent interfaces, modal dialogs, things that move when the page loads. That is where previous agent demonstrations have quietly failed, and it is not something a benchmark measures.
Watch the failure reports rather than the launch coverage, which will take a few weeks to accumulate. And watch the price, because a price set at launch is a hypothesis about value that the market either confirms or corrects.
Our earlier reporting on agents operating real machines is in this piece, and the competing release from Anthropic is covered in our report on Fable and Mythos.
Published in The Outspoken Digest
Editorial desk
Outspoken Digest Technology DeskSoftware, hardware, artificial intelligence and what they change for everyone else.
Newsletter
The Digest, in your inbox
One edition, sent when it is ready. No noise, and your address is never passed on.
Read Next
More Technology →
Anthropic Shipped the Same Model Twice Under Two Names, and the Difference Is Who Is Allowed to Use It
Sep 1, 2026/4 min read
Chrome Deleted the Last of the Old Extensions Yesterday, and the Ad Blockers Are the Point
Sep 1, 2026/3 min read
OpenAI Is Buying Mac Minis by the Ten Thousand, Because Teaching a Model to Use a Computer Needs Computers
Sep 1, 2026/3 min read