AnalysisAI ModelsSeptember 4, 2026

FrontierMath Erdős benchmark: Astra solves 2 of 68, others fail

Epoch AI's new benchmark, FrontierMath Erdős, features 68 unsolved problems selected by mathematician Thomas Bloom. A pre-release Astra solved 2, while GPT-5.6 Sol, GPT-5.5, Claude Fable 5.1, and Claude Fable 5 all failed.

1 source

More stories today

Open the live feed