OpenAI launches GPT-6 Astra

The part worth reading twice: Astra is the first OpenAI model rated Critical for cyber capability under its Preparedness Framework, which the company disclosed Tuesday. The version shipping today "will refuse to comply with more advanced cybersecurity tasks such as creating proof-of-concept exploits," while allowing secure code review and patching.1 That is close to the line Anthropic drew for Fable 5.1, which "can now be used to discover software vulnerabilities—though not to develop exploits for them."4 OpenAI says it will loosen its rules "in the coming weeks" through its Daybreak program and warns that its safety checks "can sometimes slow, pause, or stop legitimate work, including defensive cybersecurity."1

What the system card admits: "GPT-6 Astra's monitorability has decreased relative to GPT-5.6 Sol." The model is better at controlling its own written reasoning. Under adversarial testing it "can sometimes evade our internal monitors" on sabotage tasks. If the trend continues, OpenAI writes, it "seems likely" the company "would soon have significantly reduced confidence in detecting many forms of misaligned behaviors."2 Chief scientist Jakub Pachocki said the company would "withhold scaling until we can regain enough confidence," NBC News reported.5

The question we asked about Fable 5.1, answered here: Tuesday's Anthropic story ended with whether the government had signed off on its looser cyber classifiers. For Astra, the White House reviewed the model under its voluntary program before release, CEO Sam Altman told Axios on Wednesday, according to NBC News. Brockman said the reviewers asked for no substantial changes to the safeguards.5 Anthropic's page still does not say either way.

What we still don't know: which organizations have Astra today, when the wider rollout lands and what cache reads cost. OpenAI says Astra "is less likely to refuse harmless requests" than its predecessors but publishes a chart, not a rate, so the over-refusal figure Anthropic put a number on remains unquantified here.2 And every cyber and alignment result comes from OpenAI's own testing, without production safeguards, on benchmarks it built.