82.5% success rate on theorem proofs with AI h1; 68.75% across all six

Meta’s FAIR research team and New York University researchers released a paper on October 25 detailing the performance of their new automated theorem proving system, h1. The model achieved an 82.5% single-run success rate on five common Erdős problems, with an average API cost of $121 per successful proof. This performance outperformed the AlphaProof Nexus system, which recorded a 76% success rate and a higher cost of $317 per proof on the same benchmark set. However, h1’s overall success rate dropped to 68.75% when evaluated across a broader set of six problems, indicating potential scalability challenges. The study highlights significant cost-efficiency advantages for h1 compared to existing models, though its accuracy varies depending on the problem set complexity. Future developments will likely focus on maintaining high success rates across broader benchmarks while managing computational costs.