New ‘First Proof’ exam reveals how AI-powered math is a pay-to-play sport

Artificial intelligence models are advancing in mathematical reasoning, with proprietary models from major tech companies outperforming more accessible open-weight models. The First Proof project, which evaluates AI models’ math skills, highlights this disparity. Proprietary models from “frontier labs” like OpenAI are solving complex problems, such as the Navier-Stokes problem, but remain inaccessible for direct comparison. First Proof’s recent evaluation excluded OpenAI’s latest model due to its lack of public availability. Instead, they tested models like OpenAI’s GPT 6 Astra and Anthropic’s Claude Opus 5.5. The top performer, ProofCouncil, solved nine out of ten problems. This underscores the growing gap between corporate and open-weight AI models in research mathematics. QUESTION: How might the increasing reliance on proprietary AI models for solving complex mathematical problems impact the future of open-access research and innovation?