OpenAI’s Math Model: 100+ Solved Problems and an Advisory Group Without Brakes

OpenAI is reporting a record that stands out even in an industry used to superlatives: an internal model that only began training on August 28 has reportedly solved more than 100 long-standing open problems across most areas of mathematics. The company has published neither the list nor the proofs. Instead, it is pairing its announcement with an independent advisory group meant to help with the release. Together, the two show how far machine mathematics has come, and how unprepared the research community is for it.

Key takeaways

• According to OpenAI, the same internal model that reportedly solved the Navier-Stokes Millennium Prize problem in early September has since tackled more than 100 other open problems. None of this has been independently verified so far.
• Nine prominent mathematicians, including Fields medalists Timothy Gowers and Martin Hairer and physicist Edward Witten, have formed an advisory group at the Institute for Advanced Study in Princeton.
• The group works unpaid and can publish its recommendations. It explicitly does not advise OpenAI on the pace of its research.
• The trigger was an open letter from more than two dozen Fields medalists who see the mass production of AI proofs as a threat to their discipline.

What OpenAI is claiming

The September 21 announcement is brief. OpenAI writes that besides Navier-Stokes, the model has resolved “more than 100 long-standing open problems” across most areas of mathematics, and that its pace surprised the company’s own mathematicians. The post does not say which problems. OpenAI also gives no name or release date for the model. The company says it is still working out how to deploy math-related AI capabilities more broadly in a responsible way.

OpenAI described how such a result comes about when it published the Navier-Stokes proof on September 8. A system of coordinated agents, powered by the internal model, worked in groups on different versions of the problem. The successful group involved roughly 10,000 simultaneously active agents. The solution came about 88 hours after launch, and formalizing the proof in the proof language Lean with GPT-6 Astra took another 17 hours. Across all the problems attempted in that project, the agents exchanged 4.9 million messages and generated about 300 billion output tokens. In other words, this kind of mathematical research no longer happens in a chat window; it runs on industrial-scale compute.

The Navier-Stokes proof had already sparked controversy. New York mathematician Tristan Buckmaster accused OpenAI of pressure and intimidation; he had been working on a related question with an Anthropic employee. According to Scientific American, OpenAI also told the New York Times that it had made significant progress on another Millennium Prize problem, without saying which one. Reports that it is the Hodge conjecture remain unconfirmed.

Why mathematicians aren’t cheering

The criticism from the field is not aimed at the technology itself. The open letter “A Severe Misalignment of AI in Mathematics,” published September 11, acknowledges that language models can now solve major open problems. The signatories, who include Terence Tao, Peter Scholze and Maryna Viazovska, see the problem in the incentives: for AI companies, solved problems are proof of performance. For mathematics, they are only a tool on the way to the real goal, which is understanding. When statements are checked off as true or false at high speed, there is no time to work out new methods, write them up properly and credit earlier work by others. The letter explicitly criticizes that such solutions are often announced in a rush.

There is also a very practical question: who checks 100 proofs? A single major result often keeps experts busy for months. If one company delivers dozens of them within a few weeks, peer review cannot keep up, and doctoral students may lose their dissertation topics overnight.

An advisory group with a clear limit

The new Advisory Group on Mathematics and Artificial Intelligence is meant to mediate. According to its own website, OpenAI approached some of the members about setting up an external advisory board. The researchers chose instead to form an independent group at the Institute for Advanced Study and invited other colleagues to join. Besides Gowers, Hairer and Witten, the nine founding members include Ulrike Tillmann of Oxford, Ravi Vakil of Stanford and Melanie Matchett Wood of Harvard.

The group is supposed to help OpenAI assess the significance of new results and coordinate their release. Its members take no money, may offer advice nobody asked for and can make it public. OpenAI draws the line itself: the group “will not be responsible for advising us on how to pace our internal progress on mathematics.” The group, for its part, stresses that it has no decision-making power and that responsibility rests with the company. It describes its current task as coordinating the release of a large number of results that OpenAI reports its model has produced, and it is asking the math community for input through a form.

It is probably no coincidence that Gowers, a member who deliberately did not sign the open letter, is on board. In a blog post on September 17, he wrote that he agrees with much of the letter and also sees a crisis, but fears that mathematics could split into hostile camps. Seen that way, the advisory group is also an attempt to keep the conversation between the field and the industry open.

What this makes possible

For all the justified skepticism, the opportunity deserves a look. When machines not only find proofs but also formalize them in proof languages like Lean, a computer checks every single step. That is exactly what makes the flood manageable: a formalized proof can be verified by machine instead of being reviewed by hand over months. Anthropic showed how far this goes in early September, when Claude agents translated the entire proof of Fermat’s Last Theorem into Lean. For researchers, that points to a tool that takes over routine work, catches errors in published proofs and suggests new approaches that a human can then understand and build on.

The impact would reach beyond mathematics, too. The Navier-Stokes equations, for example, describe how liquids and gases move and underpin weather forecasts and aircraft design. OpenAI justifies its caution precisely on the grounds that new mathematical insights can enable applications far beyond the field.

Outlook

What matters now is whether OpenAI presents its results in a form the research community can check: with a list, a writeup and a formal proof for every single problem. Until then, “more than 100 solved problems” is a company claim, not a finding. The advisory group can bring order to the release, but it cannot slow it down. That leaves the real question with the company: whether it sets the pace of its math push by what research can absorb or by the next race against its competitors.