GPT-6 Astra will run more of your computer work on its own. That is a capability, not a permission slip.

What changed

OpenAI started rolling out GPT-6 Astra on 3 September 2026, initially to a limited set of organisations and, over the following days, to ChatGPT Plus, Pro, Business and Enterprise users, the OpenAI API and AWS. It is aimed at research, coding, computer use, document, spreadsheet and presentation work and long-running agentic tasks, and OpenAI describes it as able to do anything you can do on a computer. The API model id is gpt-6-astra, with a 1,050,000-token context window, up to 128,000 output tokens, a knowledge cutoff of 30 April 2026, text and image input with text output, and adjustable reasoning effort. Its headline API pricing, $10 per million input tokens and $50 per million output tokens, matches Claude Fable 5.1, though the two differ on cache economics, where Fable 5.1 reads cache at $0.25 per million tokens against Astra's $1. OpenAI also says Astra is the first model it has rated at the Critical cybersecurity capability level under its Preparedness Framework, which is the stated reason the rollout is phased. Its benchmark scores and its most-intelligent-model framing are vendor self-reported, and the comparison OpenAI drew to Claude was against Fable 5, not the current Fable 5.1.

Why it matters

A more capable flagship at the same price reads as a reason to hand over more, and Astra is sold explicitly as doing anything you can do on a computer. But can do is a capability claim, not a judgement that the output is right or that the action was wanted. The launch that widens what the model will attempt on its own is the same one OpenAI rated Critical on cyber capability and rolled out in stages for that reason. The distance between what it will attempt and what you can accept is now wider, not narrower, and that distance is the work.

Who should care

• Teams weighing GPT-6 Astra against Claude Fable 5.1 for research, coding or computer-use work
• People handing agentic or document, spreadsheet and presentation work to a model that now acts more independently
• Security and IT leads asked to approve a model OpenAI rates Critical on cyber capability

What to do

Pick one real task in each area you would actually use it for, research with sources, a coding change, a spreadsheet or a deck, or a multi-step agentic job, and run Astra beside Claude Fable 5.1 with the same inputs and the same acceptance checks. Read the real bill, since the two share headline rates but differ on cache cost, and keep a human checkpoint on anything the model does that you could not easily undo. Do not reach for either flagship where the missing piece is your own source, judgement or sign-off rather than more reasoning.