OpenAI has released 722 manuscripts covering 372 result families generated by an unreleased frontier model — its largest single publication of machine-produced mathematics to date. The company says the batch includes solutions to "hundreds" of open questions, a figure backed by AGMAI, the independent advisory group of mathematicians assembled to help communicate the results responsibly.
The repository includes summaries of the model's reasoning, estimates of the compute used and statistics on how many problems were attempted. OpenAI says the "average result" consumed the equivalent of roughly three hours of ChatGPT Pro thinking, and that the papers are published on GitHub with protocols for revisions and citations.
The release lands in the middle of an argument. In August, about 40 mathematicians asked OpenAI not to present unpublished model output as settled research. In late September, AGMAI urged labs to publish promptly and through established academic channels, to disclose model names, prompts and compute costs, and to "refrain from treating the release of mathematical results as marketing vehicles to promote their models" — a practice it said inflicts significant harm on the mathematical community.
Data handling is part of the dispute. Tristan Buckmaster of NYU, whose work with Levent Alpöge produced blow-up results for fluid equations related to Navier–Stokes, has publicly asked whether drafts the pair kept inside OpenAI's Codex informed the model's output. OpenAI replied that its researchers and agents did not see the work before it became public and that no specific user data was accessed, while conceding that it "cannot rule out that de-identified data derived from their usage of our products helped improve our models."
The unresolved question is verification: hundreds of manuscripts, none of them peer-reviewed, published under a company's own protocol.
