Vibeleaderboard
← All Intel
Intel / article

GPT‑6.2 Astra and GPT‑6.2 Sol

Source
simonwillison.net
Date

Context

OpenAI released two models on 17 November 2026: GPT-6.2 Astra, which it describes as its frontier model for the hardest projects, and GPT-6.2 Sol, its efficient model for everyday work. In his write-up of the launch, Simon Willison leads with price. Astra now costs $8 per million input and $40 per million output tokens, 80% of what GPT-6 Astra cost at $10 and $50. Sol stays at $2 and $10, the same as the two Sol versions before it.

The bigger cut is to cached input. When an application sends a model text it has already sent recently, such as the earlier turns of a conversation, the provider can bill those repeated tokens at a lower cached rate. For Astra that rate falls from $1 to $0.40 per million tokens. Willison calls this great news for longer-running agentic sessions. A coding working through a long task sends its growing history back to the model at every step, so much of what it sends can be billed at that cached rate.

Price is the settled improvement; quality is less clear. Willison describes the early reaction as mixed. He says Astra is not scoring quite as high as Claude Fable 5.2 on various benchmarks, and that some users find its communication style confusing. He cites Theo Browne's pre-release feedback, in which a long task stalled two hours in because the model stopped to ask an internal implementation question Browne had no way to answer. Artificial Analysis's ranking points the same way on benchmarks, placing GPT-6.2 Astra second in its Intelligence Index behind only Fable 5.2. Willison's own informal test, a single prompt asking for a three.js animation of Yonkers rising as a giant mech, is one data point and ranked Astra lower than that index does: he judged neither new model's result as good as those from Fable 5.2 or Gemini 4 Carbon.

Key takeaways · AI-distilled
  • Simon Willison notes GPT-6.2 Astra costs $8/$40 per million input/output tokens, 80% of GPT-6 Astra's $10/$50 and the same as Claude Fable 5.2. Cached input drops further, to $0.40 per million from $1, which he says favors long agentic sessions.
  • GPT-6.2 Sol keeps the $2/$10 per million token price of 6.1 Sol and 6 Sol, which Willison notes is half of GPT-5.6 Sol's $4/$20.
  • On Willison's own three.js 'Yonkers mech' prompt at Ultra effort in the Codex CLI, neither Astra nor Sol matched the results he got from Claude Fable 5.2 and Gemini 4 Carbon.
  • Willison describes early reception as mixed: Astra 6.2 scores below Fable 5.2 on various benchmarks, and Theo Browne reported it blocking two hours into a long task to ask him an internal implementation detail he could not have answered.
Terms in this piece · Glossary
  • token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
  • AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
Read the source simonwillison.net
Recommended reads
Comments

Checking sign-in…

Loading comments…