That was my experience with OpenAI's Codex auto-coder thing (running o4-mini). It took 5 minutes and like 200 commands to do what Gemini 2.5 Flash (not even Pro!) did in about 30 seconds.
I see LLMs trying to do stuff that doesn't work in every AI coding thing I've tried, despite 20 pages of system prompts! (Or perhaps because of it.)