OpenProblemBench Evaluates AI on 82 Unresolved Math and Physics Problems
October 9, 2026
OpenProblemBench benchmarks AI progress on 82 unsolved problems in mathematics and theoretical physics. In evaluations, GPT-6-Astra achieved a 14.0% mean solve rate, significantly outperforming full-size open models (5.5-6.7%) and Flash models (2.4-3.7%).
HOW THIS AFFECTS YOU
●
researcherYou can use this benchmark to test if your models are moving beyond pattern matching toward genuine scientific reasoning.