Same Codex requests, different reasoning-token counts: API versus ChatGPT Pro comparison with 960 completed responses and reproducible statistics. - johnmu/codex-reasoning-token-experiment