Codex gets two new GPT-6.2 models and a faster mode, while early testers are divided over how Astra communicates
OpenAI promises cleaner code and fewer rounds of rework, but Simon Willison and Theo Browne also see weak spots in their own trials.
OpenAI on Tuesday released GPT-6.2 Astra and GPT-6.2 Sol in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu subscribers, 9to5Mac and Thurrott reported. According to OpenAI, Astra is better at making and testing changes in an existing project; early testers are said to see a neater code style, more focused tests and fewer rounds of rework. Sol is the cheaper option for everyday programming. On Cognition's FrontierCode 2.0 benchmark, which measures whether changes are good enough for maintainers to accept, Sol at maximum reasoning effort came within half a point of Astra at roughly a fifth of the cost, according to 9to5Mac.
Also new is that Ultrafast, a mode that according to OpenAI generates tokens up to 4.5 times faster but uses up more of the allowance, is now coming to the 200 dollar Pro plan, for now only with Sol. The more expensive Pro plan gets it for both models, and Enterprise and Edu workspaces once an admin turns it on. API users pay 8 dollars per million input tokens and 40 dollars per million output tokens for Astra, a fifth less than before, according to Simon Willison, and 40 cents instead of 1 dollar for cached input. Willison calls the latter a big discount for long agent sessions. Sol stays at 2 and 10 dollars.
Willison tested both models in the Codex CLI and the desktop app on a task of his own, a Three.js animation based on OpenStreetMap data, and found them weaker there than Claude Fable 5.2 and Gemini 4 Carbon. He also cites developer Theo Browne, who was allowed to test the model in advance: Astra steers many subagents well on long tasks, but sometimes asks questions about obscure implementation details. On one task via /goal it stayed stuck for two hours on a question about naming a file. Initial reactions are therefore mixed, according to Willison.
Transparency: this edition was compiled by Claude, made by Anthropic.