OpenAI said this week that an unreleased internal model has produced solutions or substantial progress on more than 370 long-standing open problems in mathematics and theoretical computer science, a claim that has split the field between excitement over a powerful new research tool and alarm over how the results were released.
The company posted the results to a public GitHub repository on October 6 and 7, with news organizations reporting counts ranging from 370 to 377 depending on how partial results are tallied. OpenAI said nearly all of the solutions came from a single prompt to a single AI agent, each requiring an average of roughly three hours of computing time — a sharp contrast with the company's claim weeks earlier that a 10,000-agent "swarm," costing millions of dollars in compute, had produced a proof of the Navier-Stokes equations, one of mathematics' seven Clay Millennium Prize problems.
Among the new results: a claimed solution to the four-dimensional Kakeya conjecture, improvements to several well-known algorithms, and partial progress on three further Millennium Prize problems, including the Riemann hypothesis, none of which OpenAI says it fully resolved. Many of the proofs have been checked by Lean, software that verifies the logical steps of a formal argument, which mathematicians say makes the results very likely correct — though not necessarily novel or important.
A Fast-Moving, Unverified Frontier
The release follows a rockier September, when OpenAI's Navier-Stokes claim drew accusations that its model may have absorbed mathematicians' own in-progress work through its Codex coding product. OpenAI has denied this, saying its training data predates that usage. An independent advisory group based at the Institute for Advanced Study, which includes Fields medalists Timothy Gowers and Edward Witten, recommended that OpenAI publish full prompts, exact compute time and failed attempts for every claimed result, treating each like an ordinary research paper. OpenAI has disclosed average compute time and some attempt statistics but not full prompts, saying it is not formally bound by the recommendations but is trying to comply.
We should ask for receipts.
Andrew Sutherland, MIT mathematician
Sutherland's comment reflects a broader demand that single-agent claims be treated as unverified until independently replicated. NYU mathematician Tristan Buckmaster, who had separately been working on Navier-Stokes himself, put it more bluntly: "I don't think they've done their sort of due diligence at all."
Mathematicians Divided
Not every reaction has been critical. University of Toronto mathematician Daniel Litt called the GitHub release "great for mathematics," said there was no reason to keep such results secret, and predicted the tools would make researchers substantially more productive. But even Litt has warned that a public perception that AI has "solved math" could choke funding and discourage students from entering the field.
Other prominent figures are more pointed. Terence Tao has criticized what he calls the "insane" pace at which frontier labs are generating claimed results, and in September joined twenty-five Fields medalists — including Peter Scholze and Maryna Viazovska — in a signed statement warning that the race to claim famous problems as AI benchmarks is eroding the norms of verification, attribution and the human transmission of ideas on which mathematics depends. The statement argues that results are "often announced in a rush, leaving no time for a proper writeup," and insists mathematicians, not AI labs, must retain the ability to set their own research questions.
OpenAI says it plans to fund workshops and conferences to help the mathematical community evaluate the results and will release additional formal proofs over time. Even sympathetic mathematicians caution that determining whether the proofs contain genuinely new ideas — rather than clever recombinations of known techniques — could take months of human effort that the field is still organizing itself to provide.
