New OEIS Open Benchmark: Opus 4.8 Solves 30% of Unsolved Math Conjectures

xeophon · x · 2026-08-12

tmkadamcz released OEIS Open, a new mathematical benchmark comprising 492 unsolved math conjectures formalized in the Lean language.

Benchmark results show that, when provided with simple tools and a budget of $50 per conjecture, Claude Opus 4.8 can successfully generate Lean proofs to resolve 30% of the problems. This marks a significant breakthrough for large language models in advanced mathematical reasoning and formal theorem proving.

Original post →

More from Models

Models channel →