OpenAI has published 722 manuscripts detailing solutions to mathematical problems generated by an unreleased frontier model. These documents cover 372 result families and include summaries of the model's reasoning, compute estimates, and statistics on the number of attempts made. The company states that the average result required compute equivalent to three hours of ChatGPT Pro processing.
The release follows a September announcement where OpenAI claimed its model had resolved over 100 long-standing mathematical problems. The newly formed Advisory Group on Mathematics and Artificial Intelligence (AGMAI) noted that the current batch contains solutions to hundreds of open questions. The group was established to oversee the responsible communication of such findings.
AGMAI issued recommendations in late September, urging AI laboratories to publish results through formal academic channels. The advisory body specifically requested that companies disclose model names, prompts, and compute costs. Furthermore, the group advised against using mathematical breakthroughs as marketing tools to promote AI models, citing potential harm to the academic community.
OpenAI has made the results available via a GitHub repository, which includes established protocols for citations and paper revisions. The company is currently evaluating community-hosted platforms for future releases to better align with committee guidelines. OpenAI also plans to improve the quality of mathematical exposition and presentation in subsequent updates.
The mathematical community is currently evaluating these findings alongside other recent outputs from OpenAI and Anthropic. These include results related to a Millennium Prize problem. The rapid pace of these releases has sparked ongoing discussions regarding research ethics, academic conduct, and the appropriate attribution of human contributions in AI-generated mathematical research.
