Epoch AI's Mystery Game Puzzles benchmark conceals which game it's testing so models can't be tuned to the test; Opus 5 leads at 59%.
Continue to AI University →