OpenAI Rates GPT-6.2 Astra High for AI Self-Improvement
It is the first OpenAI model deployed outside the company to receive a High rating in the self-improvement category; GPT-6.2 Sol did not.

GPT-6.2 Astra is the first OpenAI model deployed outside the company to earn a High rating for self-improvement, a capability area focused on AI systems helping develop other AI systems.
OpenAI launched GPT-6.2 Astra and GPT-6.2 Sol on Nov. 17. The company’s Preparedness Framework sorts potentially dangerous capabilities into categories and assigns ratings such as High and Critical. Its self-improvement category assesses whether a model can help develop AI systems, which could speed the development of future models.
GPT-6 Astra, released in September 2026, did not meet High in self-improvement. GPT-6.2 Sol also fell short in that category, while receiving a Critical rating for cybersecurity and High ratings for biological and chemical risks.
OpenAI’s ongoing safety work includes further assessment of self-improvement.