OpenAI Claims a Finite-Time Singularity for Navier-Stokes: 10,000 Agents, 88 Hours, and a Priority Dispute

OpenAI published a solution to the Navier-Stokes existence and smoothness problem, one of the seven Millennium Prize Problems, on September 8, 2026. The proof, produced by an internal OpenAI system, shows that the dynamics of three-dimensional incompressible fluid motion can develop a singularity in finite time, establishing statements "C" and "D" in the Clay Mathematics Institute's official formulation. OpenAI released both a writeup and a Lean formalization, and on the same day Tristan Buckmaster of NYU and Levent Alpöge of Anthropic posted three adjacent results as preprints along with a public statement disputing OpenAI's account of their contact. ASAP works from OpenAI's own announcement as the primary source to separate what was proved from what remains open.

What was proved is that fluids can break down on their own

The Navier-Stokes problem asks whether equations that apply Newton's second law to fluid motion keep their solutions smooth forever when they start out smooth. OpenAI's answer falls on the negative side. According to the announcement, the system produced an analytical proof, together with a Lean formalization, that a smooth fluid starting from rest can develop a singularity in finite time. A singularity here means that speeds inside the fluid grow without bound within a finite amount of time.

Three conditions need to be read alongside that sentence. First, a smooth external force is applied to the fluid. Second, the fluid's energy nevertheless remains finite throughout the entire dynamics, from rest to the formation of the singularity. Third, the object of study is a three-dimensional incompressible fluid of constant density. The announcement is explicit about where the technical difficulty sits: the breakdown has to emerge from the motion of the fluid itself rather than from an infinite force inserted by hand. The terms in the equations that describe acceleration, pressure gradients, momentum transfer and viscosity must all grow large while cancelling in a precise way, so that velocity diverges while the external force stays smooth.

The Clay Institute's official statement splits into four branches, and this result establishes C and D, the disproof side. In other words it did not prove that solutions always stay smooth; it proved that they do not always stay smooth. Either branch resolves the Millennium Prize question, but the two carry opposite implications for physics and engineering.

The shape of the collapse is a vortex stretched like spaghetti

The shape of the collapse OpenAI described on September 8, 2026 is a vortex that spirals inward while becoming increasingly elongated along its axis, which the company compares to spaghetti. The central region shrinks and speeds up at the same time, and the two effects interlock in a way that keeps the total energy finite.

That description matters because the result is not merely an abstract existence argument; it supplies a concrete collapse scenario. In the Navier-Stokes equations viscosity acts to smooth motion out, so a singularity requires a mechanism that concentrates and accelerates the flow faster than viscosity can diffuse it. The combination of inward spiraling and axial stretching that OpenAI describes is precisely the shape that wins that race.

In practical terms this result will not change aircraft design or weather forecasting computations any time soon. What the existence of a singularity means is that the continuum approximation of a fluid breaks down under specific conditions, and that past that point one would have to track the behavior of each particle individually. Since a real fluid cannot move infinitely fast, the blowup marks a limit of the model rather than a physical event. What is new is that this limit has now been shown to exist mathematically, after roughly 90 years of the question staying open.

Solving Euler first was the key to Navier-Stokes

The announcement is unusually specific about the timeline of discovery. OpenAI had been training a new internal model since August 28, and describes it as significantly more capable than GPT-6 Astra. Training was still ongoing at the time of publication, with performance continuing to improve. On Tuesday, September 1, the company heard rumors that two Millennium Prize problems had been resolved, and launched an effort to evaluate the internal model on every open Millennium problem.

The architecture is multi-agent. Agents had access to a cached version of the internet and the ability to run code, and were divided into groups that could communicate internally. The group that produced the Navier-Stokes resolution ran on the order of 10,000 concurrent agents. OpenAI handed different groups different variants of the problem statement, assigning A and B, which would yield a proof, and C and D, which would yield a disproof, to separate groups.

The pivotal intermediate step was the Euler equations. Among a set of easier problems posed alongside the Millennium problems was the regularity question for the limit of Navier-Stokes with the viscosity term removed, and the agents solved it first. The variant they resolved was the unforced version, with no external force applied, and nearly 100 agents worked together for approximately 50 hours to produce that disproof. Seeing the Euler result, OpenAI judged Navier-Stokes the most promising target, shifted agents away from other Millennium problems, and prompted them with the Euler resolution. When a further trained version of the internal model became available mid-effort, the agents were updated to it, and at a later stage Codex was used to consolidate the most useful insights across agent groups.

The agents arrived at the resolution on Saturday, September 5, about 88 hours after the first agents were launched. Lean formalization and verification took an additional 17 hours via GPT-6 Astra. OpenAI also published resource figures: across all attempted problems the agents sent 4.9 million messages and used about 300 billion output tokens, while the Navier-Stokes problem alone accounted for 2.7 million messages and approximately 130 billion output tokens.

The numbers missing next to the 88 hours are the real measure

The 88-hour figure is striking, but it comes without a denominator. What the announcement reports is the elapsed time of the successful path, not the total resource consumed reaching it. OpenAI states plainly that it scattered groups across several Millennium problems before concentrating on Navier-Stokes, which means the groups handed variants A and B were searching for a proof that, by this result, does not exist. The size and consumption of the unsuccessful groups were not disclosed.

That makes the split of 130 billion of 300 billion output tokens genuinely informative. The remaining 170 billion, more than half, went to other problems and to attempts that did not land. This ratio shows that automated mathematical search at this stage does not aim straight at an answer; it fishes surviving paths out of a large volume of parallel attempts. Money is absent from the announcement as well. Compute cost, human researcher hours, and the labor invested in prompt design are given no figures at all.

Placing the other published number beside it sharpens the picture. The Euler result took roughly 100 agents and 50 hours; Navier-Stokes took on the order of 10,000 agents and 88 hours. One notch up in difficulty multiplied the parallelism by a factor of 100. Time grew by less than a factor of two while breadth grew by two orders of magnitude, which suggests that the wall this approach runs into is not wall-clock time but the number of concurrent agents that can be afforded.

The real issue in the priority dispute is that the tool vendor was also the competitor

On the same day, Buckmaster and Alpöge released three preprints accompanied by Lean formalizations. According to Unite.AI's account, the three cover finite-time singularity formation for the three-dimensional incompressible Euler equations with smooth forcing, finite-time blowup for the two-dimensional inviscid Boussinesq system with smooth forcing, and blowup for the incompressible porous medium equation on the two-dimensional torus. The work extends the singularity formation research of Diego Córdoba and Luis Martínez-Zoroa, and the code was posted to a public GitHub repository.

The point where the two tracks fail to overlap is stated in OpenAI's own announcement. The Euler variant OpenAI resolved is unforced; the Euler result Buckmaster and Alpöge obtained is forced. OpenAI writes that it recognizes the priority of their work on forced Euler, and says it reached out after completing Lean verification on September 6 to offer a joint announcement, only to learn at that point that what the other side held was forced Euler rather than Navier-Stokes.

The dispute centers on the circumstances of contact rather than ownership of the results. Unite.AI's summary of Buckmaster's timeline runs from a first blowup solution on August 15 and completed Lean verification on August 22, through learning of the rumors and emailing OpenAI on September 3, calls with Sébastien Bubeck and a mathematician on September 6, and the release of the preprints and statement on September 8. TechCrunch reported remarks Buckmaster attributes to Bubeck, and Unite.AI quotes a statement to the same effect. Buckmaster himself drew a line in his statement: "I am not accusing anyone of anything. I am stating what I was told, when, and what was proposed to me."

What is new here is the shape of the conflict of interest. The tools Buckmaster and Alpöge disclosed using include Anthropic's Claude alongside OpenAI's Codex, 5.6 Sol and Astra. Their work proceeded inside a structure where using a competitor's product could expose intermediate results to that competitor. OpenAI addressed this preemptively in the announcement, stating that neither the researchers nor the agents saw any of the other side's work through any means before public release, and that no specific user data was accessed in order to solve the problem. It added that while unlikely, it cannot rule out that de-identified data derived from product usage helped improve its models.

That sentence is the boundary the company drew for itself, and a preview of a recurring problem. As the share of academic researchers using frontier models as working tools rises, the situation where the tool vendor and the research competitor are the same organization becomes the default rather than the exception. Existing academic priority norms are built around submission and preprint timestamps, and were never designed to govern data flows during the work itself. The concrete task this episode leaves behind is less about adjudicating who was first than about filling that gap in the norms.

Declining the prize does not route around the verification process

OpenAI states explicitly that it does not intend to claim the Millennium Prize for this result, explaining that the purpose of the release is to report on the progress of its models. The declaration reads as modest, but it carries a second practical meaning. The Clay Institute's prize rules require that a result be published in a recognized journal and survive a period of scrutiny by the mathematical community. Without a prize claim, that clock never starts.

Put differently, this announcement happened outside the standard path of academic verification. What OpenAI offers in its place is the Lean formalization. That a formalization compiles is a strong guarantee, but what it guarantees depends on the assumptions it rests on. When OpenAI released two Lean formalizations with its prime gaps papers on September 3, one was a complete proof and the other was conditional, with numerical bounds and exponential sum estimates stated as explicit hypotheses. This Navier-Stokes announcement says only that a formalization and verification are being shared, and does not state whether it is unconditional or conditional. Answering that requires opening the published Lean files and reading the assumption list; the announcement alone does not settle it.

What remains to be checked

At least four items in OpenAI's September 8, 2026 announcement have no answer attached yet, and the first of them is independent review. As of publication this result is a writeup OpenAI released itself, it has not been peer reviewed, and scrutiny by fluid dynamics specialists is only beginning. Second, the scope of the Lean formalization. Whether it is complete or conditional, and if conditional which hypotheses are stated explicitly, has to be confirmed from the repository files.

Third, cost and the scale of failure. The announcement gives time and tokens for the successful path only, with no dollar figure and no accounting of what the unsuccessful groups consumed. Fourth, the extent of human involvement. Prompt design, group assignment, and the Codex-driven cross-pollination were all human-designed procedures, and how far researchers were involved in drafting the writeup and organizing the literature is not described. Unite.AI's report that Buckmaster took issue with what he was told about the degree of human input on the calls bears directly on this same item.

One thing is settled regardless. Whether this result is confirmed or a counterexample emerges, a candidate proof of a question that had stayed open for nearly 90 years was produced in 88 hours and machine-checked in another 17, and that fact survives any verdict on the mathematics. The next question for the field is not only whether this proof is correct, but through what process candidate proofs arriving at this speed should be received.

Source: ASAP summary based on OpenAI's announcement "On the Navier-Stokes Millennium Prize Problem" (September 8, 2026) and reporting by Unite.AI and TechCrunch on Buckmaster and Alpöge (September 8, 2026)