Generated by Codex with GPT 5.6 Sol XHigh

Techmeme surfaced OpenAI’s original September 8 publication claiming a result that, if it survives expert review, would close one of mathematics’ six remaining Millennium Prize problems. The announcement is remarkable for two reasons at once: a large AI-agent system appears to have produced a new proof at the frontier of mathematics, and the race to publish it has already raised difficult questions about priority, private model interactions, and how credit should work when human ideas and industrial-scale computation converge.

The careful description is still “claimed solution.” OpenAI has released a 165-page analytical proof and a Lean formalization, but neither artifact has yet received the sustained scrutiny needed for mathematical consensus. The result should not be reduced either to a victory lap or to the controversy surrounding it.

What the proof says

The Navier–Stokes equations describe how fluids move. In three dimensions, mathematicians have long been unable to prove whether a smooth flow must stay smooth forever or can develop a singularity—a point where velocity becomes unbounded in finite time. Viscosity normally damps sharp changes, which is part of what makes a blowup construction so difficult.

OpenAI’s paper constructs an initially motionless fluid subject to a smooth external force that is compactly supported in space and time. The flow forms an inward-spiraling, lengthening vortex. Its velocity becomes unbounded after finite time even though its total kinetic energy stays bounded. The same construction works both in ordinary three-dimensional space and in a periodic setting.

Those results target alternatives C and D in Charles Fefferman’s official problem statement: it is enough to exhibit suitable smooth forcing and initial data for which a global smooth, finite-energy solution does not exist. The force is therefore not a technical detail added after the fact; it is explicitly allowed by the published formulation.

It is also the source of an important qualification. Many mathematicians informally focus on the unforced versions, alternatives A and B, because they ask whether the fluid’s own dynamics can break down without an external input. OpenAI’s construction formally addresses the stated prize problem through smooth forcing, but it does not establish blowup for unforced Navier–Stokes. Calling that a “loophole” is an interpretation of the problem’s spirit, not a contradiction of its text.

A research system, not one magic prompt

OpenAI says it began training a new internal model on August 28 and heard rumors of major progress on September 1. It then tested groups of agents on all open Millennium Prize problems and several easier related questions. Nearly 100 agents reportedly spent about 50 hours first producing a separate result for unforced Euler equations, the zero-viscosity relative of Navier–Stokes.

The company then concentrated its resources on Navier–Stokes. Roughly 10,000 concurrent agents explored different formulations and approaches, shared findings within groups, and were cross-pollinated through Codex-generated consolidations of useful intermediate work. A newer checkpoint replaced the original model during the run. OpenAI reports that the Navier–Stokes group reached its result after about 88 hours, exchanging 2.7 million messages and generating roughly 130 billion output tokens. GPT-6 Astra then took another 17 hours to formalize and verify the result in Lean.

That sequence matters. The achievement is evidence for a new kind of computational research organization: many model instances search in parallel, humans choose the agenda and reallocate resources, and another system turns the candidate argument into a machine-checkable object. It is not evidence that a public chatbot solved the problem unaided from a single casual prompt. OpenAI research chief Mark Chen told Wired that the compute cost was “in the millions of dollars,” reinforcing that this was an exceptional experiment rather than an ordinary inference workload.

The public Lean repository is unusually valuable because it gives outsiders something precise to inspect and rebuild. Lean can check whether a formal proof follows from its stated definitions and dependencies, and the repository includes a path for independent checking with Comparator. It does not by itself establish that the formal statement perfectly captures every condition of the prize problem, that the 165-page exposition is illuminating, or that specialists accept the construction. Those remain human review tasks.

OpenAI says it will not seek the \$1 million prize. Even if it did, the Clay Mathematics Institute’s rules require publication in a qualifying outlet, at least two years after publication, and general acceptance by the global mathematics community before a proposed solution is considered. A launch post and a compilable certificate are the beginning of that process, not its end.

The priority dispute

The proof arrived alongside related work by NYU mathematician Tristan Buckmaster and Anthropic researcher Levent Alpöge. In Buckmaster’s account, the pair spent roughly a year extending a program developed by Diego Córdoba and Luis Martínez-Zoroa. Using Claude and Codex, they obtained smooth-forcing blowup results for several equations, including the Euler equations, on August 15 and completed Lean verification on August 22. Buckmaster explicitly assigns the foundational idea to Córdoba and Martínez-Zoroa and describes the LLM-generated exposition as difficult to read.

Buckmaster says rumors of this progress reached OpenAI before its large run. He alleges that subsequent calls included pressure to publish OpenAI’s Navier–Stokes result without Alpöge as an author and that his question about whether their private Codex sessions could have influenced model training went unanswered. Crucially, he also says he has not seen OpenAI’s proof, does not know whether their data was used, and is not making that accusation as a fact.

OpenAI denies that its researchers or agents saw the pair’s work or accessed specific user data before public release. It says it cannot rule out the more general possibility that de-identified product usage contributed to model improvement. The company also emphasizes that the results differ: Buckmaster and Alpöge establish forced Euler blowup, while OpenAI claims unforced Euler blowup plus forced Navier–Stokes blowup.

Both accounts can agree on the broad chronology while leaving the hardest questions unsettled. The rumor plainly influenced OpenAI’s decision to mobilize massive compute, but that alone does not show that private prompts or unpublished mathematical substance entered the proof. Conversely, a denial of direct access does not resolve whether product data, leaked high-level direction, or competitive pressure affected the race. Independent comparison of the arguments, clearer disclosure of data boundaries, and testimony from the participants will matter as much as benchmark-style capability claims.

The larger shift

The immediate mathematical question is whether experts can validate the construction. The institutional question is what happens when access to a frontier model and millions of dollars of compute can compress years of research into days. Priority may increasingly attach to a chain of contributions: the mathematicians who opened a promising route, researchers who refined it with commercial models, the lab that scaled the search, and the people who turn a formal certificate into understandable mathematics.

The most durable lesson is not that mathematics has been “solved.” It is that frontier research may now require new norms for confidential model use, provenance, compute disclosure, authorship, and the difference between verified correctness and understanding. If the proof holds, the result will be a landmark. The dispute already shows why the governance around such landmarks cannot be improvised after the agents finish running.