On Codex, the GPT-5.5 model with over 200K context can’t handle execution at all. Every time I research a big task, it has to auto-compress. Also, in my test scenarios (Swift), Claude is still better at writing code.