OpenAI began rolling out GPT-6 Astra on September 3, describing it as "the world's most intelligent and aligned model" and the product of years of work across pre-training, reinforcement learning and alignment. President Greg Brockman called the release a "generational leap" that could eventually be seen as an early step toward artificial general intelligence.

Benchmark claims and a new safety threshold

OpenAI says Astra is state-of-the-art on computer use, web browsing, software engineering, cybersecurity and scientific reasoning. The company cites a 98% score on FrontierMath Tier 4, 99.9% on ARC-AGI-3 and a perfect 100% on ExploitBench, its internal exploit-development benchmark. On Terminal-Bench 4.0, a coding-agent benchmark, OpenAI reports Astra scoring 57.7%, ahead of its own GPT-5.6 Sol (37.3%) and Anthropic's Claude Fable 5.1 (55.8%).

That cybersecurity performance comes with a notable caveat: Astra is the first OpenAI model designated as reaching the "critical" risk threshold for cyber capability under the company's preparedness framework, meaning it can identify and exploit previously unknown vulnerabilities in well-protected systems without step-by-step human direction. OpenAI says it has applied additional safeguards and access restrictions in response to that classification, though specifics of those mitigations were not fully detailed in the launch materials.

Codex gets a long-session memory upgrade

Alongside Astra, OpenAI is updating its Codex coding environment with an experimental feature that lets the model take running notes across multiple context windows during long sessions, instead of compressing everything into a single summary each time a window fills up. Earlier context remains searchable, so Astra can look up requirements or test results from earlier in a session even if they weren't captured in its notes. OpenAI says the feature will become the default in Codex in the coming weeks, and credits the underlying harness changes with a 1.9x faster task-completion rate.

Rollout and why it matters

Astra began rolling out first to enterprise customers with "Daybreak" access on September 3, with availability to ChatGPT Plus, Pro, Business and Enterprise users, the OpenAI API and Amazon Web Services following in the coming days. The launch pushes OpenAI's flagship model line and its Codex coding agent forward together, and the "critical" cybersecurity designation is one of the first times a major lab has publicly flagged a shipped model as capable of unsupervised offensive cyber work — a milestone likely to sharpen scrutiny of how frontier labs test and gate increasingly capable coding and computer-use agents before release.