OpenAI Says Its AI Solved 100 Unsolved Math Problems

OpenAI made two announcements on Sunday that only make sense read together. The first is a capability claim: an internal model has resolved more than 100 open mathematical problems, and the pace of that progress, in the company's words, has surprised the mathematicians inside OpenAI. The second is an institution: a nine-member advisory group of prominent mathematicians assembled to review and communicate what the model produces. A lab that had simply solved hard problems would publish and let the field respond. OpenAI is building the response mechanism at the same time as the results, which tells you where the real bottleneck now sits.

The claim is 100 problems, and the proof of it is the problem

OpenAI says the model resolved more than 100 open problems across most areas of mathematics, and it points to a claimed solution of Navier-Stokes existence and smoothness, one of the seven Millennium Prize Problems the Clay Mathematics Institute has posted at a million dollars each since 2000. That is an extraordinary claim on its own. The company has not released the proofs for peer review, and Santage could not independently verify any of the specific results beyond OpenAI's description of them.

Verification is the whole of mathematics as a working discipline. A proof becomes a result only when other mathematicians read it, follow every step, and agree the argument holds. That process runs on human time. A referee report on a single hard paper can take a year, and a Millennium Prize claim would draw scrutiny measured in years. A model that generates candidate proofs faster than any group of humans can check them produces a growing backlog of claims waiting on a verification layer that has never scaled, rather than a set of settled questions.

A nine-member advisory group is really a verification layer

The advisory group is the tell. OpenAI named nine mathematicians with weight the field respects, drawn from the Institute for Advanced Study, Cambridge, Stanford, Berkeley, Oxford, and beyond. Their stated role is to advise on the review and communication of emerging results and to act as a bridge to the mathematical community, while keeping editorial independence from OpenAI. The members are unpaid, can publicize their own views, and control their own membership.

What the group explicitly does not do is advise OpenAI on how fast to push its internal mathematics work. The pacing stays with the lab, and the outside experts are positioned around the output. Read plainly, OpenAI has recruited credibility rather than oversight. The names lend the results standing with the public and the press before the slow machinery of peer review has said anything, and the group's independence is what makes that borrowed standing worth having.

The 25 Fields Medalists who signed against the race

The field's own reception is divided. An open letter signed by 25 Fields Medal winners argues that AI labs are threatening mathematicians' intellectual work as they compete to be first with solutions to famous problems. Only one of OpenAI's nine advisers, Camillo De Lellis, also signed that letter, which places the advisory group and the field's most decorated objectors on largely different sides of the same question.

The tension is real and worth naming. Mathematicians spend careers on single problems, and the value of that work is partly the understanding built along the way, not only the answer at the end. A model that outputs answers without the human path to them can look, from inside the discipline, less like a collaborator and more like a machine that skips the part that gave the work meaning. The letter is a signal that a large share of the field will not grant automatic legitimacy to lab-produced proofs, advisory group or not.

Why the bottleneck moved from capability to review

For anyone tracking where AI is heading, the shift here matters more than any single proof. The frontier labs spent the past year showing that models can take on research-grade work, and the GPT-6 Astra generation pushed reasoning far enough that a claim of 100 solved problems is at least plausible rather than absurd. The constraint is no longer whether a model can generate a candidate proof. It is whether the human systems that certify knowledge can keep up with the volume, and whether they will choose to.

OpenAI's move is a bet that the answer is institutional. Build the credibility structure, seat respected names around the output, and the verification gap becomes manageable rather than disqualifying. Whether that works depends on people the company does not employ and cannot direct, which is exactly why it needed the group and exactly why the group had to be independent to be worth anything. The hardest problem OpenAI's model surfaced this week sits outside mathematics entirely. It is the question of who gets to certify that a machine has proved something, and how long the rest of us should wait before believing it.