OpenAI · Model release
GPT-6 Astra: OpenAI's computer-use flagship arrives gated
OpenAI's first GPT-6-family model is a computer-use flagship — and the first model it gates at the Preparedness Framework's 'Critical' cybersecurity threshold.
What changed
- Trained on >100,000 GPUs at Stargate — OpenAI's largest training run to date — and rolled out over several days from 2026-09-03 across ChatGPT, the API, Azure, and Bedrock.
- Priced $10/$50 per MTok up to 272K input and $20/$75 above it; cached input is $1 short or $2 long. Fast costs 2x applicable token rates. Ultrafast lists $60/$300 short and $120/$450 long on Responses with global/US processing; actual latency depends on the workload.
- Benchmarks: 74.1% (xhigh) on Datacurve's independent DeepSWE 1.1 capture dated 2026-09-22 at $4.43/task; the captured board labels it top among listed configurations, not a current universal crown. OpenAI reports 72.6% on OSWorld 2.0 offline; the dated 2026-09-26 rebased AA Intelligence Index v4.3.2 snapshot shows 53 (max, publisher-rounded).
Best fit
- Computer-use agents
- Agentic coding
- Hard reasoning with vision
What to be careful about
- Gated at the 'Critical' cyber threshold; Enterprise tenants get it disabled by default and eligibility checks apply.
- AA Intelligence Index scores are specific to a dated version and effort setting; do not infer a universal ranking from the v4.3.2 snapshot.
- OSWorld figures are OpenAI's own; the DeepSWE score now matches Datacurve's independent run.
Dataset snapshot: GPT-6 Astra is also represented in the checked comparison dataset. Prices and benchmark figures carry their own field dates.