First Proof Benchmark Exposes Wide Gap Between Proprietary and Open AI Math Skills
The First Proof project released its third benchmark results, revealing that proprietary AI models like Anthropic’s Claude Opus 5.5 and OpenAI’s GPT 6 Astra are now capable of solving complex research-level math problems, while open-weight models lag significantly behind. A combined system called ProofCouncil solved nine of ten rigorous problems, marking a major leap from earlier rounds. However, the leading open-weight model, Kimi K3, failed to solve any of the ten problems, highlighting a substantial performance gap. This disparity has intensified concerns among mathematicians about data privacy and the accessibility of advanced AI tools, as open-source alternatives remain inadequate for high-level research. The results follow a turbulent period in AI-math, including OpenAI’s controversial mass release of unverified proofs and Google Deepmind’s peer-reviewed breakthroughs.