OpenAI released two new models on November 17: GPT-6.2 Astra for the hard projects and GPT-6.2 Sol for everyday work. Both are live now in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users. In the API they’re called gpt-6.2-astra and gpt-6.2-sol.
What OpenAI is promising
Astra is supposed to build slides, spreadsheets and documents that stick more closely to templates and the conventions of a field. OpenAI says it can match a slide template down to the font size, even from a screenshot. On Agents’ Last Exam, a benchmark for multistep office work, Astra scores 5.2 points above GPT-6 Astra and 5.5 points ahead of Claude Fable 5.2, at 5 percent of Fable’s cost. A footnote says Fable 5.2’s result leaves out nine tasks where a fallback model stepped in.
Harvey LAB-AA only counts a legal task when every requirement is met and nothing is made up. There, Astra reportedly more than doubles Fable 5.2’s pass rate at about a sixth of the cost.
Sol is the workhorse. On Cognition’s FrontierCode 2.0 coding benchmark, it lands 0.5 points behind Astra at roughly a fifth of the cost.
Pricing and the cyber threshold
In the API, Astra costs $8 per million input tokens and $40 for output, 20 percent less than GPT-6 Astra. Cached input is $0.40. Sol stays at $2 and $10.
Both models hit the “Critical” cybersecurity threshold in OpenAI’s Preparedness Framework, as GPT-6 Astra did before them. So OpenAI is launching them with tighter filters on risky security requests, and says itself that this also slows down legitimate vulnerability research. Vetted security researchers are supposed to get the new models later through its Daybreak program.
What the independent numbers say
Every chart in the announcement comes from OpenAI. The Decoder put Artificial Analysis’s numbers next to them. In the Intelligence Index, which rolls up nine evaluations, GPT-6.2 Astra comes second, just behind Claude Fable 5.2. Gemini 4 Carbon follows with 60 points and Claude Opus 5.5 with 59. Sol reaches 57, which matches GPT-6 Astra at about 15 percent of the cost per task.
Same list price, different appetite
Claude Fable 5.2 has cost $8 input and $40 output per million tokens since November 9. Per token, the two models cost exactly the same. So when OpenAI talks about 5 percent or a sixth of the cost, the gap comes from consumption: Astra supposedly needs far fewer tokens for the same task. Those cost figures come from OpenAI’s own runs, and the only way to know if they hold for your work is to compare both on real tasks. For Anthropic, that means list price is no longer a gap anyone can argue about.