OpenAI’s latest mathematical reveal arrived after weeks of anticipation: 722 manuscripts spanning 372 families of related results, produced by an unreleased frontier model and said to address “hundreds” of open questions.¹ A separate report framed the disclosure around findings on 377 math problems, underscoring how quickly the field is trying to sort and verify the scale of the claims.²

The company had previewed the moment in September, saying its system had resolved more than 100 long-standing problems across most areas of mathematics. The new release adds paper drafts, some reasoning summaries, compute estimates and statistics on attempted problems. OpenAI says the average result required the equivalent of three hours of ChatGPT Pro thinking.¹

Its public message is unapologetically expansive. In announcing the batch, OpenAI said it had consulted the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study.Lilian Weng (lilianweng)³ President Greg Brockman cast the effort as part of a push “towards acceleration of scientific discovery and improving quality of life for everyone.”Greg Brockman (gdb)⁴

But the advisory group’s own recommendations draw a sharper boundary. Published in late September, they urged AI labs to disclose mathematical results promptly, preferably through established academic channels, and to reveal the model name, prompts and compute costs.¹ Most pointedly, the group said companies should “refrain from treating the release of mathematical results as marketing vehicles to promote their models,” warning that the practice can harm the mathematical community.¹

OpenAI says it is publishing through a GitHub repository with procedures for revisions and citations. That offers a route for scrutiny—but not necessarily the peer-review process critics want. The real test now is less whether the model found results than whether the results can survive the slower, stricter machinery of mathematics.