Claude Mythos solves OpenAI's Erdős problem with a different proof approach, Anthropic engineer reports

Anthropic engineer Sholto Douglas wrote on X that Claude Mythos can solve OpenAI's recently disproved Erdős unit-distance conjecture—an open problem in combinatorial geometry since 1946—with a 'cute, simple proof,' indicating what he called 'serious overhang' in AI-driven math discoveries.

The team tested Mythos using isolated Claude Code instances that receive the problem, develop solution paths independently, and then distribute findings across further instances. Mythos frequently took a different mathematical route than OpenAI's model, though mathematician Daniel Litt called the result 'a bit worse' than OpenAI's; Mythos reportedly found OpenAI's solution as well.

Google DeepMind separately announced that an AI-assisted system solved nine Erdős problems using the formal proof language Lean, an approach noted as less impressive from a pure language model perspective compared to systems like Claude Code, which functions as an agentic harness (a system that can plan and execute tasks autonomously).