ChatGPT 5.2 Thinking, Gemini 3 Pro, and Claude Opus 4.5 all beat human top scorers at Tokyo and Kyoto University — two years after GPT-4 failed to pass.