GPT-6 Astra

OpenAI's first GPT-6-generation model, launched September 3, 2026 as `gpt-6-astra` after a month of staged safety disclosure: previewed August 1 through ten solved open mathematics problems, then confirmed on September 1 as the first model OpenAI has ever classified **Critical** for cybersecurity under its Preparedness Framework (it can find and weaponise zero-days in hardened targets without human direction). Rolling out first to enterprises in the Trusted Access Program, with API, Plus, Pro, Business and Enterprise access following; advanced offensive-cyber capability stays gated behind Daybreak Blue. $10/$50 per Mtok (cached input $1.00, cache writes $12.50; >272K-token prompts billed at 2x input / 1.5x output; Fast mode 2x at up to 2.5x speed), 1,050,000-token context, 128K max output, April 30 2026 knowledge cutoff. PROVENANCE CAVEATS — every recorded row is OpenAI-run at maximum effort, and OpenAI does not say whether standard users can select that effort. Two headline claims are deliberately NOT recorded here: the ARC-AGI-3 99.9% (run through a Responses-API harness that retains reasoning across turns and compacts context — OpenAI has itself shown those settings can triple ARC-AGI-3 scores without changing the model, so the number measures the agent system; the public ARC Prize leaderboard's best result is Claude Opus 5 at ~30.2%, and the viral 98.6%/99.9% figures were flagged as unverified by third-party coverage) and the FrontierMath Tier 4 98% saturation claim (no independent corroboration). Its DeepSWE lead is also narrower than the launch chart suggests: OpenAI's own chart omits Muse Spark 1.3, which reports 75.4% against Astra's 74.1%.