2026-09-05 GPT-6 Astra release

gpt-6-astra — Saturated benchmarks, Critical cyber rating

OpenAI's GPT-6 Astra saturates ARC-AGI-3 (99.9%), FrontierMath Tier 4 (97.6%), and ExploitBench (100%) in one release, while becoming the first OpenAI model to meet the Critical threshold for cybersecurity under its Preparedness Framework — it found two real zero-days on its own during evaluation. Astra is out and its good — almost unreal looking at benchmarks. I've only just started playing with it.

See also OpenAI's companion safety writeup, Path to Astra: critical capabilities and frontier safeguards (Sep 1, 2026): Astra wasn't involved in the Hugging Face incident, but training was paused for two weeks afterward to harden isolation controls; Astra refuses 91.5% of cyber-jailbreak attempts vs. 59% for GPT-5.6 Sol.