OpenAI publishes 722 math papers, solving 90 of top 500 open problems

OpenAI published a large batch of mathematics results from an internal frontier model in a public GitHub repository, totaling 722 manuscripts grouped into 372 related result families. The batch came from an evaluation of roughly 4,000 research problems, and reports say it solved 90 of 500 top open mathematics problems. The release includes papers, proof artifacts, and partial reasoning summaries, but the model itself remains unreleased.

Two sharply different narratives have emerged around the same set of materials. Sam Altman called it 'a new era of discovery'; the mathematician Levent Alpöge praised the quasi-Riemann and no-Siegel-zeros results, calling them 'the most important moment in the history of mathematics.' But the core result numbers and evaluations come mostly from OpenAI's own blog, repository, and supporters' comments, and the results have not yet been independently verified.

Among this batch of results, Result 003, the 'quasi-Riemann hypothesis,' was singled out and described as falling between a Fields Medal–level achievement and 'the greatest result in number theory in 200 years.' Commenters also pointed to partial progress on the Riemann, Hodge, and BSD problems. These evaluations come from individual commenters and have not yet been independently verified.

A distinction is needed: the publication of the manuscripts and the family classification are confirmed facts, whereas the academic status of the 'quasi-Riemann hypothesis' and the partial progress on the Millennium Problems currently remain source claims. Whether they truly reach the heights claimed depends on subsequent independent review.

Reports say each result consumed an average of about three hours of ChatGPT Pro thinking compute. By comparison, the same source mentions that Navier-Stokes took 88 hours and deployed 10,000 agents. This contrast is used to emphasize the 'efficiency' of the new results.

But Teortaxes pointed out that three hours of compute 'is not much,' which suggests another reading: low compute could indicate an efficient method, or it could indicate that the task itself did not touch the hardest parts. Whether three hours of compute is enough to indicate the difficulty and reproducibility of the results remains an open question.

One analysis estimates that about 20% of the results are disproofs or counterexamples, and argues that this weakens the claim that 'AI mathematics victories rely mainly on brute-force search.' This proportion comes from one analysis, not an official statistic.

If that estimate holds, it does point to a capability different from exhaustive enumeration; but disproofs and counterexamples likewise need to withstand scrutiny. Whether the roughly 20% proportion is accurate, and how strongly it supports the 'non-brute-force search' argument, cannot currently be confirmed.

The skepticism is not aimed at the release itself, but at its extrapolation. Will Depue expects some results to fail scrutiny, and for that reason built citedbyagi.com to track which human papers the release cites. François Chollet, meanwhile, questions whether gains on mathematics and code suited to RLVR can generalize, or whether unverifiable domains remain limited by human data.

These two doubts point to the same gap: even if this batch of mathematics results holds up, it does not automatically follow that the model has comparable capability in broader domains. The generalization problem and the scrutiny problem are the two weakest links in the current chain of evidence.

OpenAI says it consulted the Institute for Advanced Study's independent advisory group on mathematics and artificial intelligence about how to release these results. This move indicates that the publisher is aware of the sensitivity of the results and the need for verification.

But consultation is not the same as verification. The model was not released, the results have not been independently reviewed, and Alpöge's mention of related reports about other lab users being scooped and about conflicts of interest means the credibility assessment of this batch cannot rest solely on the publisher's narrative. The specific circumstances and scope of the scooping and conflicts of interest remain unclear.