A 99-problem benchmark drawn from 117 papers and patents; GPT-6.2 Astra scores 84.5%.
GPT-6.2 Astra is the frontier model for the hardest projects and Sol the efficient everyday model; Astra API prices drop 20% and both are available in ChatGPT, Codex and the API.
Both models target agentic coding and code review; they roll out gradually to Pro+, Max, Business and Enterprise.