Codex is also much less transparent about its reasoning. With Claude, you see a fairly detailed chain-of-thought, so you can intervene early if you notice the model veering in the wrong direction or going in circles.