GPT-6 Astra

What it is

GPT-6 Astra is OpenAI’s new model generation, rolling out from 3 September 2026. It is a generation change, the first since GPT-5, and it takes the top seat in OpenAI’s range from GPT-5.6, whose three tiers stay on sale beneath it. The API lists it at $10 per million input tokens and $50 per million output, with cached input at $1, a 1,050,000-token context window that accepts roughly 922,000 tokens of input, and up to 128,000 tokens of output. That pricing puts it level with Anthropic’s Claude Fable 5.1, which shipped two days earlier at the same $10 and $50.

The Critical rating

The most consequential line in the launch material is a safety classification. GPT-6 Astra is the first model OpenAI classifies Critical for cybersecurity under its Preparedness Framework, in the deployment safety report it published on 3 September. A Critical rating is OpenAI’s own statement that the model’s offensive-security capability crosses the highest threshold the framework defines, and it arrives with deployment controls to match. The rating puts OpenAI alongside Google, whose Gemini 3.8 Flash Cyber ships only to vetted defenders through the Fairwind Program: the frontier labs now treat top-end cyber capability as something to be gated, and they say so in public documents.

Where it lands

On the independent boards the debut is strong. The rebased Artificial Analysis Intelligence Index of 5 September 2026 scores GPT-6 Astra at maximum effort on 55, second behind Claude Fable 5.1 on 57 and ahead of Claude Opus 5 on 54; it had not yet entered the top ten of the LMArena text board at the same check. The two results YFarmX has covered draw the capability picture at its edges: ARC Prize scored it at 62.7% and 99.9% on ARC-AGI-3 on the same day, depending on which protocol you count, and the Astra name first surfaced alongside ten formal proofs with machine-checkable Lean certificates.

What to watch

Three things will settle where this generation stands. The first is independent testing beyond the launch-week boards, on the long-horizon agentic work the price implies. The second is how the Critical cyber classification shapes availability, and whether capability withheld from the general release surfaces through gated programmes the way Google’s Fairwind does. The third is the answer from the rest of the field: Anthropic reset the top of the index two days before Astra arrived, and Google’s next Pro-tier move is still in partner testing. Our large language models hub tracks all of it.