This is my feedback after using both model families on 23 November 2025.
I recently had a very productive session with Codex while refactoring tests and adding new features. I used both Codex-max and Claude, but to be honest, Codex-max on high settings works much better for me.
The usage limits at OpenAI are also much more generous. I think Codex with the Plus plan for 20 is a much better value than Claude Max for 100.
Codex is a better model for people who already have programming experience. It follows instructions very closely, so you must be careful what you ask for. But if you know what you want, it feels like a very good pair-programming partner.
Claude models right now feel more like tools for “vibing.” Maybe they work better for non-technical people. For me, they do not follow instructions so well, even after asking many times. Claude often breaks the logic or refuses to fix the code properly, even when the task is very clear.
Because the quality of models changes fast, I think it is important for a company to stay open and let developers choose the best tool for their work. This helps avoid a kind of “natural selection” where one tool is forced on everyone.
There are also other good and cheap models, like GLM (I still need to test it more).