OpenAI Says an Internal Model Solved Navier–Stokes — and Stands Up an Outside Math Advisory Group

OpenAI says that an internal model it began training on August 28 has resolved the Navier–Stokes Millennium Prize problem and more than 100 other long-standing open problems across most areas of mathematics. In the same post, the company says it is working with mathematicians who have set up an independent advisory group, hosted at the Institute for Advanced Study, to help review and communicate results and to advise on how OpenAI's tools should serve mathematical research. The announcement contains an extraordinary capability claim and almost none of the evidence a mathematician would normally expect alongside it.

What OpenAI actually said

The post is short on specifics and long on process. OpenAI states that the model is internal, that training started on a specific date, and that the pace of progress "surprised the mathematicians within OpenAI." The company does not name the model, describe its architecture or scale, explain how the results were produced, or say whether any proof has been written up, submitted, refereed or checked by a formal proof assistant. No timeline for publication or external verification is given. No independent mathematician is quoted confirming any single result.

That gap matters more here than in a typical product launch. "Resolved" is doing heavy lifting: in mathematics, a claimed resolution becomes a resolution only after other experts read the argument and agree it holds. The Navier–Stokes existence and smoothness problem is one of the Clay Mathematics Institute's Millennium Prize problems, and the standard route to recognition runs through publication and sustained scrutiny, not a corporate blog post. OpenAI's own framing implicitly acknowledges this by saying the new advisory group will help it assess the significance of emerging results and coordinate their dissemination — which reads as an admission that the assessment has not happened yet.

The "more than 100 long-standing open problems" figure is similarly unqualified. The post does not define what counts as long-standing, list the problems, or indicate how difficulty was judged. Open problems in mathematics range from questions a strong graduate student could close in a week to questions that have defeated the field for a century. Without a list, the number is not interpretable.

The advisory group, and what it is not

OpenAI names nine initial members: François Charles (ENS-PSL), Camillo De Lellis (IAS, GSSI), Timothy Gowers (Collège de France, Cambridge), Martin Hairer (EPFL, Imperial College London), Nikhil Srivastava (Berkeley, Simons Institute), Ulrike Tillmann (Oxford, INI), Ravi Vakil (Stanford), Edward Witten (IAS) and Melanie Matchett Wood (Harvard). The group is hosted at the Institute for Advanced Study.

The governance terms OpenAI describes are unusually clean for a corporate advisory body. Members are not paid by OpenAI. The group controls its own membership. It can offer advice OpenAI did not ask for, comment publicly on OpenAI's impact on mathematics, and publish its advice. OpenAI says the group's value depends on members exercising independent judgment and challenging the company — a line that at least sets an expectation the group can be measured against.

The limits are stated just as plainly. The group will not advise OpenAI on how to pace its internal progress in mathematics. That carve-out is the most consequential sentence in the announcement. It means the body can weigh in on how results are described, verified and shared, and on how tools reach researchers, but not on whether or how fast the underlying work proceeds. Advisory groups without a say over the thing that generates the risk tend to end up ratifying decisions rather than shaping them. Whether this one is different will depend on how its members use the public-comment freedom they have been given.

There is also no published charter detail in the post: no cadence of meetings, no term lengths, no description of what information OpenAI will share with the group or when, and no mechanism for the group to see results before they are announced. Those are the operational details that determine whether a bridge to the community functions or is decorative.

The complaint that prompted this

OpenAI points to an open letter titled "A Severe Misalignment of AI in Mathematics," in which mathematicians raise concerns about using open problems as benchmarks for new AI systems and the negative externalities that follow. The post does not summarize the letter's arguments in detail or name its signatories.

The underlying tension is familiar to anyone who has watched benchmark culture in machine learning. When a hard, legible target becomes a marketing milestone, incentives shift toward claiming the target rather than doing the slower work of verification and exposition. Mathematics has specific vulnerabilities here: refereeing capacity is finite and volunteered, and a flood of machine-generated arguments could consume reviewer time that has no substitute. There is also a credit and attribution question the post does not address — who is named as author on a result produced by a company's internal model, and what happens to a researcher who spent years on a problem that a model closes without notice.

Why product teams should care

For people building on frontier models, the interesting signal is not the Millennium Prize claim, which they cannot evaluate, but the fact that OpenAI says it is "working through how broader deployment of math-related AI capabilities can be done responsibly." That is language typically used ahead of staged or restricted release. The post gives no availability date, no pricing, no API surface and no indication of whether any of this capability will reach developers at all. Anyone planning around it is planning around a research claim.

The second signal is structural. Advanced mathematical reasoning, if real and general, tends to travel: optimization, verification, cryptanalysis and scientific modeling all sit near it. OpenAI gestures at this by calling mathematics a fundamental science with wide application. But capability in proving theorems does not automatically transfer to engineering utility, and the post offers no evidence on transfer either way.

The honest summary is that OpenAI has made a claim that would be among the most significant results in modern mathematics, published no proof, and simultaneously created an outside body whose first job appears to be figuring out how such claims should be handled. Those two things in one post are in tension. The advisory group's first real test is whether it can put verified detail behind the Navier–Stokes claim — or say publicly that it cannot.

What to watch

• Whether a preprint, formalization or referee-ready writeup of the Navier–Stokes claim appears, and how long the gap is between the announcement and any checkable artifact.
• Whether the advisory group publishes anything on its own initiative, including advice OpenAI did not request — the clearest test of the independence OpenAI describes.
• Whether OpenAI releases a list of the "more than 100" problems it says were resolved, with enough specificity for outside mathematicians to evaluate difficulty.
• Any move from internal-only to external availability of math capabilities, and what access tiers or restrictions come with it. None were disclosed.
• How the Clay Mathematics Institute and the broader community respond, and whether journals or funders adopt any standard for AI-generated proofs as a result.