My takeaway is that the “winner” depends on which layer you evaluate. Claude Code may win on speed and immediate product usability, while Codex may win on architecture, testing and reliability. I’d score them separately on time to first usable version, regression coverage, failure handling, maintainability after several changes, and total cost. Giving both tools the same prompt is a useful baseline, but it may not represent the workflow in which each agent performs best.