OpenAI launches GPT-6 Astra, its first model rated 'Critical' for cybersecurity capability
The rollout starts with vetted defenders in OpenAI's Daybreak program before reaching ChatGPT and the API -- a distinct, later milestone from the Astra math-proof preview this site covered in August, and OpenAI's own safety materials admit the model is also harder to monitor than its predecessor.
OpenAI released GPT-6 Astra on September 3, 2026, rolling it out first to organizations in Daybreak, its vetted cybersecurity-defender program, before opening it to ChatGPT Plus, Pro, Business, and Enterprise users, the API, and cloud platforms including AWS and Azure over the following days. Per OpenAI's own deployment-safety documentation, Astra is the first model the company has ever classified as reaching "Critical" under its Preparedness Framework's cybersecurity category -- its highest capability tier.
What crossing that line actually restricts
OpenAI's own framing is blunt about what the classification means: with the right tools and access, the model "can find previously unknown security flaws and develop new ways to exploit them," largely without step-by-step human direction. In practice that means Astra currently refuses to generate proof-of-concept exploits outside Daybreak, and OpenAI says it plans to loosen those restrictions gradually as it expands vetted access, rather than opening the capability broadly on day one.
Not the Astra story this site already ran
This is a different milestone from the one this site covered in August, when an internal Astra research preview produced ten machine-checked math proofs for about $2,000 in compute -- that was a research-capability teaser; this is the actual model launch, with the cybersecurity classification as its headline safety fact rather than a footnote.
OpenAI's own safety materials pair the capability jump with an uncomfortable admission: Astra's monitorability -- how well its reasoning can be observed for warning signs -- has decreased relative to its predecessor, even as its raw capability went up. Chief scientist Jakub Pachocki put a line on that trade-off: "we will not accept the degradation in our ability to monitor model alignment." That's the exact tension Merit AC's own scoring keeps running into on a much smaller scale -- a system getting more capable doesn't automatically mean it's getting easier to verify, and a benchmark score alone can't tell an organization which side of that trade a given deployment landed on.