The Pulse
OpenAI Says Its Internal Model Solved More Than 100 Math Problems
OpenAI says a new internal model has resolved more than 100 long-standing open problems across mathematics, alongside its claimed Navier–Stokes result. The company is forming an independent advisory group to help review and communicate the
AI.info Team ·
OpenAI says a new internal model has resolved more than 100 long-standing open problems across mathematics, but the company has not released a list of those results or supporting papers. The announcement arrives as mathematicians question whether AI companies are turning difficult problems into publicity tests before the work can be checked and understood.
The claim appears in an OpenAI post published September 21, 2026, announcing an independent advisory group on mathematics and artificial intelligence. OpenAI says the model began training on August 28 and has also produced what the company describes as a resolution of the Navier–Stokes Millennium Prize problem.
The company presents the advisory group as a response to the pace of progress and to criticism from mathematicians. The group will review the significance of emerging results, advise on their release, and comment on academic and professional standards. It will operate independently, receive no payment from OpenAI, and have the freedom to publish advice or challenge the company.
OpenAI’s claim arrives without a catalogue of proofs
OpenAI’s announcement gives no breakdown of the more than 100 problems. It does not identify the problems by name, state which fields they cover, provide papers for each result, or say which have been reviewed by outside mathematicians. The wording also leaves open how OpenAI defines “resolved,” a distinction that matters in research mathematics, where a plausible argument, a formal verification and community acceptance are different stages.
The company’s earlier account of its Navier–Stokes effort offers more detail about the internal system. OpenAI said its agents sent 4.9 million messages and generated about 300 billion output tokens while working across the problems it attempted. The group focused on Navier–Stokes after finding what it described as promising results on a related Euler-equation problem, then used Lean formalization and verification through GPT-6 Astra.
OpenAI’s Navier–Stokes post says the agents reached their resolution on September 5, with formalization and verification taking another 17 hours. The company also says it does not intend to claim the $1 million Millennium Prize, which is awarded by the Clay Mathematics Institute under its own review process.
Mathematicians object to speed without understanding
The announcement follows a September 11 declaration signed by dozens of Fields Medalists and other mathematicians. The declaration argues that solving open problems as an AI benchmark can harm mathematics when companies announce results before the field has time to examine the proof, isolate new ideas, credit earlier work and explain the result to other researchers.
Signatories distinguish between establishing whether a statement is true and developing the understanding that makes a result useful to mathematics. They warn that a large supply of machine-generated true-or-false answers could weaken the human processes through which mathematical ideas are tested, taught and incorporated into the discipline.
That dispute gives OpenAI’s advisory group a narrow but significant assignment. The group is not being asked to control the company’s internal research schedule. OpenAI’s announcement explicitly says the advisers will not determine how quickly the company pursues mathematical progress. Their role is instead to assess results, shape their communication and advise on how the work should reach researchers and the public.
Independent in structure, limited in authority
The group is hosted at the Institute for Advanced Study and includes François Charles, Camillo De Lellis, Timothy Gowers, Martin Hairer, Nikhil Srivastava, Ulrike Tillmann, Ravi Vakil, Edward Witten and Melanie Matchett Wood. Their public website says the group formed after OpenAI approached some of its members about an external board. The mathematicians chose an independent structure and say they can offer advice to any AI company whose systems have a significant effect on mathematics.
The group’s website also describes its immediate task in unusually direct terms: advising OpenAI on how to coordinate the release of a large number of significant mathematical results that the company says its internal model produced. That wording confirms the group is dealing with OpenAI’s account of the results, not announcing that it has independently validated all of them.
OpenAI’s own post says the advisers will help assess significance and academic standards, but it does not say that they have verified the more-than-100 figure. Until the underlying problems, proofs and review status are published, the number is best treated as an OpenAI claim rather than an established tally of accepted mathematical results.
The next test is publication, not scale
OpenAI has already published a separate set of ten mathematical and theoretical-computer-science advances produced by an internal version of Astra. That earlier release named each problem, linked to a paper and reasoning walkthroughs, and said the arguments were formalized into Lean certificates. The September 21 announcement provides none of that documentation for the larger set.
The advisory group may help determine whether those materials arrive as a single release, a series of papers or a process that allows mathematicians to examine results before public claims are made. Its members can publish recommendations, but they do not have decision-making power inside OpenAI. Responsibility for the claims and the evidence will stay with the company.
For now, the concrete record is a company statement dated September 21: OpenAI says an internal model trained from August 28 has resolved more than 100 long-standing open problems. The proofs, problem list and outside assessments needed to judge that statement have not yet been released.
Read OpenAI’s announcement, the advisory group’s public charter, and the mathematicians’ declaration.