GPT-6 Astra Is Live — OpenAI's Flagship At Apex
September 5, 2026
GPT-6 Astra is in the model selector today. It is OpenAI's most capable model for the hardest end-to-end work — complex reasoning, coding, research, and long agent loops — with a 1.05M-token context window and 128K max output. It sits at quality 71, in the Apex 18× band.
Independent boards already put it at the front of the pack. ARC Prize has Astra at 95% on ARC-AGI-2 and 62.7% on ARC-AGI-3 with their standard harness — both state-of-the-art on those boards. Artificial Analysis ranks GPT-6 Astra (max) #2 on the Intelligence Index.
What you can do
- Pick GPT-6 Astra from Closed SOTA in the model selector. It is the highest-rated model in the catalog.
- Keep the whole project in one context — 1.05M input and 128K output, with tools and high-effort reasoning on by default.
Where this shows up
You have a multi-day refactor or a research brief that already outgrew GPT-5.6 Sol. Pin GPT-6 Astra and keep going in the same thread.
You want Anthropic's Mythos-class pass and you know Fable 5.1 does not support zero data retention. Pin Fable yourself. For the same Apex 18× band without that constraint, pick Astra.
Try it
- "Use GPT-6 Astra to find the root cause of this rare crash — disassemble if you need to, then write the fix and the tests."
- "Challenge this market-entry plan, research the missing evidence, and produce a board-ready recommendation."
- Pin GPT-6 Astra, then: "Try again — go deeper on the assumptions that do not hold."
Heads up
- Note (September 9): Auto and Swarm no longer select Astra. Pin it when the turn deserves Apex — it is too expensive to auto-pick. Fable 5.1 stays out of both pools because it is non-ZDR; Astra stays out because of cost. Astra is still ZDR.
- Apex 18× — the same token band as Claude Fable 5.1. Reach for it when the turn deserves the top of the catalog, not as an everyday default.
- Community accounts on the Budget 1× ceiling still cannot reach Apex until review approval or a purchase. Paid plans can pin it.