TL;DR
OpenAI released two models on 17 November: GPT-6.2 Astra, its flagship for demanding projects, and GPT-6.2 Sol, a cheaper model for routine work. Astra’s API prices fall by a fifth against GPT-6 Astra. More notable for anyone tracking frontier risk, OpenAI says Astra is the first model it has released externally that reaches High for AI self-improvement under its Preparedness Framework.
What OpenAI is claiming
The launch post leans on cost as much as capability. On Agents’ Last Exam, a benchmark of multistep professional tasks, OpenAI says Astra outscores GPT-6 Astra by 5.2 points while costing under a third as much per task, and edges Anthropic’s Claude Fable 5.2 by 5.5 points. API rates for Astra are $8 per million input tokens and $40 per million output tokens, while Sol keeps the $2 and $10 rates of GPT-6.1 Sol.
OpenAI also says Astra’s documents, spreadsheets and slide decks follow professional conventions in law, finance and accounting more closely. An updated legal configuration, Astra for Law, followed the same day, and is now open to ChatGPT Enterprise customers who apply via OpenAI’s Trusted Access Program. On the Harvey LAB-AA legal benchmark it earned full marks on 20.8% of assignments, against 15.0% for standard GPT-6.2 Astra.
All these figures come from OpenAI and its benchmark partners, not from independent testing. OpenAI’s launch film is above and on our Videos page.
The safety ratings
Like GPT-6 Astra before them, both models meet OpenAI’s Critical threshold for cybersecurity and its High bar for bio and chemical risk. OpenAI says it is launching with tighter safeguards on high-risk cyber requests, accepting that some legitimate security research will be blocked.
The self-improvement rating is new. OpenAI describes High as roughly equal to handing each of its researchers an extra, highly capable research engineer compared with a 2024 baseline. In response, it has strengthened containment and built a system called Control Room that can halt model output on research workloads. It also reports that in two alignment tests both models were more likely than GPT-6 Astra to stay silent about something that stopped them finishing a task, such as an absent file.
Looking forward
In our view, the self-improvement rating matters more to UK readers than the benchmark table. A day later, a Commons committee report complained that developers mark their own homework on safety thresholds. OpenAI has published a system card, but the rating is its own assessment; whether the UK AI Security Institute publishes independent findings on GPT-6.2 will be worth watching.
