OpenAI’s pitch is one sentence long: anything you can do on a computer, Astra can do for you. The benchmark table underneath it is more interesting than the slogan.
GPT-6 Astra landed on Thursday evening, and OpenAI is not selling it as a better chatbot. It is selling it as something that sits at your keyboard. The company calls it state of the art on computer use, browsing, software engineering, cybersecurity and scientific work, and the demo framing is that Astra reads pixels and drives a mouse the way a person does, rather than calling an API someone had to build first.
That distinction is the whole product. An agent that needs a bespoke integration per application only ever works with software that shipped one. An agent that can look at a window and click things works with KiCad, FreeCAD, Power BI, a Python notebook, or whatever ancient internal tool your company refuses to replace. OpenAI lists all of those, plus filling forms, updating CRM records and installing and troubleshooting software.
The numbers, and what they actually say #
On an offline subset of OSWorld 2.0, Astra scores 72.6% against 65.7% for GPT-5.6 Sol. The more striking figure is time: roughly 40 minutes per task versus about 75, which OpenAI puts at around 47% less. Elsewhere the company claims 98.6% on ARC-AGI-3, 97.6% on FrontierMath Tier 4 v2, 74.1% on DeepSWE v1.1 and 100% on ExploitBench.
Read that time figure twice, though. Forty minutes per task is a very long time to watch a cursor move by itself, and it is a reminder that “superhuman speed” in the marketing refers to individual actions, not to finishing the job. The OSWorld result is also an offline subset rather than the full live benchmark, which is a meaningful asterisk on any computer-use score. And a 100% on an exploit benchmark says as much about the benchmark being saturated as it does about the model.
The part that should make you sit up #
Astra is the first model OpenAI has classified as Critical for cybersecurity under its Preparedness Framework, the top rung. In the company’s own description that means it can find and build working zero-day exploits against hardened real-world systems without human guidance, and can plan and run novel end-to-end attacks given only a high-level goal.
OpenAI is gating the full cyber capability rather than shipping it to everyone: alpha access goes to people defending critical infrastructure, including the US government and its Daybreak partners. It also reports that Astra refused 91.5% of prohibited cyber requests against 59% for GPT-5.6 Sol. That is a real improvement and still leaves roughly one in twelve getting through, on a model the vendor itself rates as capable of serious harm.
Price, availability, and the bit nobody mentions #
API pricing is $10 per million input tokens and $50 per million output, with a Fast mode running up to 2.5 times quicker at double the rate. Enterprises in the Trusted Access Program get it now; Plus, Pro, Business and Enterprise users, the API, AWS Bedrock and Azure follow over the coming days. OpenAI is pushing customers to think in price per completed task rather than per token, which is a convenient framing when your per-token number is that high, and a fair one when the alternative is a human doing it.
Worth remembering what running this on your own machine implies. We covered in August that ChatGPT’s computer-use mode keeps a log of every click and keystroke it makes on a Mac. Handing a model that much reach over a desktop is a genuine convenience and a genuine attack surface, and the “AGI era” language the launch has attracted does nothing to help anyone reason about which is which.
Sources: OpenAI’s announcement post and Path to Astra, with reporting from VentureBeat and The New Stack.