The same can be said about Claude (no or tiny Opus on Pro) vs GPT-5.2 high (5.3-codex if you like terminal bench hacking).
Anecdotally GPT was also smarter than Claude which prompted my move from Claude in the first place: Gemini and Claude back in October failed to get their own harness PID.
Outside of anecdata I rely on https://artificialanalysis.ai/models/capabilities/coding for now.