OpenAI releases 722 math manuscripts from unreleased model

OpenAI published 722 mathematical manuscripts generated by an unreleased internal artificial intelligence model on October 6, 2026. The works are organized into 372 clusters of related results and were shared on GitHub alongside materials designed to help researchers inspect their proofs.

The company evaluated roughly 4,000 problems during the project. Because the 372 groups include main findings, consequences, and alternative proofs, the 722 manuscripts do not represent 722 distinct solved problems. OpenAI noted that the findings remain in varying stages of verification, and some unformalized papers may contain errors.

Part of the proofs were formalized in Lean, a programming language and proof assistant that checks mathematical arguments by computer. OpenAI stated that it will update the repository with new formalizations as they become available, keep previous versions accessible, and record corrections. The release also includes ten reasoning summaries, with each result requiring computing resources equivalent to roughly three hours of ChatGPT Pro reasoning on average.

OpenAI consulted the independent Mathematics and AI Advisory Group at the Institute for Advanced Study before releasing the archive. In recommendations published on September 29, the advisory panel urged AI labs to publish documentation, process details, and formal proofs, while requesting funding for community-led research. However, the group emphasized that its consultation is not an endorsement, urged labs to stop benchmarking advanced mathematics on closed proprietary models, and noted that machine-generated proofs do not replace human comprehension.