I've been using gpt-5 on effort=high but for gpt-5-codex, try: `-c model_reasoning_effort=medium`.

On high it is totally unusable.

even on medium ... gpt-5 was way faster, at least that's my first impression