OpenAI has published 722 mathematics papers produced by an internal frontier model that it has not yet released, organized under 372 result families, on GitHub. According to the independent AGMAI group, the works contain results addressing hundreds of long-unsolved open problems; however, they still need to be verified by the mathematics community and some may contain errors.
The model encountered approximately 4 thousand problems during evaluation; an average result was produced with roughly three hours of ChatGPT Pro thinking capacity. OpenAI shared reasoning summaries for 10 problems:
- the irrationality measure of the number π
- the Mahler conjectures
- Kaplansky’s direct finiteness conjecture
- the three-dimensional relativistic Vlasov-Maxwell system
A portion of the proofs is being verified with the Lean system; the repository will continue to be updated. One work concerning the Riemann zeta function, however, was produced through an exceptional process and edited by a human. The repository also includes rules for updating the papers and for providing citations.
AGMAI, based at the Institute for Advanced Study in Princeton, wants the results to be published through academic channels, transparency, and for the discoveries not to be turned into a marketing tool.
Background
OpenAI is not a new name in the FikirPilot archive: we have published 80 stories mentioning this name in the last 90 days; the most recent is dated October 6, 2026.