What OpenAI shipped on November 17
OpenAI released GPT-6.2 Astra and GPT-6.2 Sol for harder projects and everyday work. The API names are gpt-6.2-astra and gpt-6.2-sol. OpenAI says both are in ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu users. The November 18 briefing carries the release with its source.
Standard API prices, per million tokens, are the ones OpenAI published and that Artificial Analysis restates: Astra at $8 input, $0.40 cached input, and $40 output, about 20 percent below GPT-6 Astra. Sol stays at $2 input, $0.10 cached input, and $10 output, the same standard rates as GPT-6.1 Sol. These are list prices. They are not a CuratedAI measurement of your bill.
A lower Astra price is not a coding upgrade
Artificial Analysis, in the same November 17 write-up, puts GPT-6.2 Astra second on its Intelligence Index, 5 points above GPT-6 Astra and behind Claude Fable 5.2. That jump puts Astra above Gemini 4 Carbon at 60 and Claude Opus 5.5 at 59. GPT-6.2 Sol gains 2 points, to 57, which Artificial Analysis says is level with the previous GPT-6 Astra. Those are their scores, at the effort settings they name.
The coding result on that page points the other way. On the Coding Agent Index, run with the Codex harness at maximum effort, Artificial Analysis says Astra is flat with GPT-6 Astra, and Sol loses 3 points against GPT-6.1 Sol. The Sol drop is driven by a 14-point fall on SWE-Atlas-QnA-AA, where the model more often left out facts the test expected. On AA-Omniscience, the same write-up reports hallucination rates at maximum effort of 59 percent for Astra and 55 percent for Sol. That is a tendency to guess on their factual set, not an error rate inside your product.
Match the model to the cost of being wrong
Sol is one-fourth the token price of Astra at the standard rates above. That is the reason to start there. Astra is the model to try when the independent index moved and a miss is expensive. Neither score tells you which one finishes your queue.
Price the finished task, not the token
- Replay a sample you are allowed to keep. Include ordinary requests, long prompts, and one case that previously failed. Save the model id and settings with each run.
- Score the thing the user needed. A parsed field, a cited passage, or a change that still builds. Count a fluent answer that drops a condition as a miss.
- Record tokens and retries. Cached input only helps when the same prefix is actually cached. A cheaper token still loses if the model spends more output, or if a person has to repair it.
- Keep the two models on the same tools. A harness change can look like a model change. The model migration test is the longer version of this check.
Where the release is easy to overread
Second on an intelligence index is a rank across someone else's exams. It does not move your default. Sol remaining at $2 does not mean it is the same model as GPT-6.1 Sol: the coding index moved down even though the list price did not. ChatGPT Work and Codex are the surfaces OpenAI named. Check the picker in the product you have before you tell a team the default changed.
Reviewed November 19, 2026, from OpenAI's introduction post, the November 18 briefing, and Artificial Analysis's November 17 index note. For the neighboring models already in the directory, see GPT-6 Astra and GPT-6.1 Sol.
Questions
- When should you use GPT-6.2 Astra instead of GPT-6.2 Sol?
- Use Sol first when the job is repeated and checkable. Try Astra when a missed detail is expensive, then keep whichever model finishes your own sample at an acceptable cost.