Atlas

Benchmarks

← All benchmarks

FrontierMath Erdős

Math · 2026-09-03

68 formal conjectures covering 65 difficult Erdős problems open as of August 2026. An Inspect DeepAgent with offline mathematics literature attempts a Lean proof or disproof once per conjecture, within $300 and 72 hours. Comparator verifies submitted proofs in an isolated container.

Top models (higher is better)

ModelScore
GPT-6.1 Sol4.4
GPT-6 Astra2.9
Claude Opus 5.52.9
Claude Sonnet 5.52.9
GPT-6 Sol1.5
Claude Fable 5.10.0
Claude Fable 50.0
Claude Opus 50.0
Gemini 3.8 Flash0.0
GPT-5.50.0
GPT-5.6 Sol0.0
Grok 4.70.0
Kimi K30.0
Loading Atlas data…