OpenAI launches GPT-6 Astra, rates it Critical for cybersecurity

SAN FRANCISCO — OpenAI released GPT-6 Astra on Thursday, Sept. 3, 2026, calling it the company’s most capable model yet for computer use, browsing, software engineering, science, and cybersecurity, and opening a limited Daybreak rollout that expands to ChatGPT Plus, Pro, Business, and Enterprise users and the API over the coming days, TechCrunch and OpenAI’s own posts reported.

President Greg Brockman told reporters Astra is OpenAI’s “most intelligent and, also very importantly, our most aligned model yet,” and closed the briefing with “Welcome to the AGI era,” while declining to treat AGI as a contractual trigger after OpenAI’s Microsoft deal changed. The company says Astra is state of the art on computer-use and coding benchmarks and the first model it has designated at the Critical cybersecurity threshold under its Preparedness Framework — meaning, with the right tools, it can find previously unknown flaws and build exploits across hardened systems without step-by-step human guidance, per OpenAI’s Path to Astra safety post.

That cyber rating is why the launch is staged. Advanced cybersecurity workflows start with Daybreak testers and Daybreak Blue defenders; production access still layers refusals, classifiers, and monitoring that can pause agent runs when the system flags misuse or unauthorized actions. OpenAI says Astra refused 91.5% of cyber jailbreak prompts in its tests versus 59% for GPT-5.6 Sol, and that retrospective testing suggests current production safeguards would have blocked the earlier Hugging Face incident — though Astra itself was not the model in that breach. Critics and reporters also flagged “opaque recurrence,” a reasoning style that can make chain-of-thought monitoring harder as models do more work with fewer visible tokens.

For users, the pitch is agentic work: multistep computer tasks, polished documents and sites, and stronger coding in real repos. For everyone else, the story is the usual concentrated-power problem dressed as a product launch — a private lab shipping a model it rates Critical for cyber offense, then asking the public to trust phased access and auto-review pauses. Astra is live for a limited set of organizations today; Plus and API users are next in line.