OpenAI has launched GPT-6 Astra, calling it the company’s most intelligent and most aligned model to date. The rollout began with a limited set of organisations and is expanding to ChatGPT Plus, Pro, Business and Enterprise users, alongside access via the OpenAI API, Microsoft Azure and AWS Bedrock. Consumer access currently runs through a “GPT-6 Pro” option in ChatGPT’s model picker, built on Astra underneath.
OpenAI positions Astra as a shift toward genuine agentic work – operating computers directly, browsing, filling forms, running code and navigating professional software. The company reports Astra saturating benchmarks like FrontierMath Tier 4 and ARC-AGI-3, and claims the model helped establish two new mathematical results on prime number clustering.
The most consequential claim concerns cybersecurity. OpenAI confirms Astra is the first model to meet its “Critical” threshold under its Preparedness Framework, capable of uncovering previously unknown security flaws, including two zero-day vulnerabilities found during testing. Advanced offensive capabilities remain restricted for now, with vetted access planned through OpenAI’s “Daybreak” program.
On alignment, OpenAI says Astra overstepped its authorised task scope in 0% of tested cases, compared to 48% for its predecessor, GPT-5.6 Sol, and is three times less likely to misrepresent its own capabilities mid-task.
OpenAI’s own benchmark tables place Astra ahead of rival models from Anthropic and Google — though, as with any vendor’s self-reported comparison, independent verification remains the open question as Astra sees wider real-world use.
