OpenAI says it has used AI to solve one of the Millennium Prize Problems. On 8 September 2026, the company announced a proof concerning Navier–Stokes, the equations we use to describe fluid motion. It claims to have constructed a flow that starts without irregularities and develops a singularity: velocities growing without bound in finite time.
If the proof is confirmed, it would be an enormous advance on a question that has remained open for decades. The announcement comes with a dispute involving mathematicians Tristan Buckmaster and Levent Alpöge over priority, authorship and research data. But even setting that controversy aside, there is a mathematical reason not to consider the matter closed.
The construction uses an external force. Showing that a fluid can lose regularity with that intervention does not, by itself, decide whether it can do so without it.
That detail matters because the prize statement permits this route. We could have a solution to the Millennium Prize Problem while still lacking an answer to one of the most natural questions about the equations.
OpenAI’s announcement links to a manuscript and a formalization in Lean, a proof assistant. Here we examine the scope of the claim; neither the announcement nor the existence of a formalization is presented as a substitute for independent mathematical scrutiny.
What it means for a fluid to lose regularity
Water flowing through a pipe and air moving around an aircraft are examples of flows studied with Navier–Stokes. We have used these equations for over a century. Their usefulness does not require knowing everything they can do in every imaginable situation.
The Clay problem concerns a specific version: an incompressible fluid, with positive viscosity, in three dimensions. Incompressible means that each small portion keeps its volume as it moves. It can deform and rotate, but it cannot be compressed.
We start with a smooth velocity field, without jumps or initial singularities. We know how to follow the solution for at least some time. The question is whether we can continue indefinitely or whether the evolution itself can break that smoothness. This is called blow-up.
We are not seeking a formula to calculate every possible movement of water. We are asking whether the smooth mathematical description can cease to exist in finite time.
Two tendencies coexist within the equations. Viscosity smooths velocity differences and dissipates energy. Nonlinear dynamics allow motion to reorganize its own structure; in three dimensions, vortex stretching can intensify rotation and generate smaller scales.
The intuitive picture is a competition between smoothing and concentration. Saying that viscosity smooths does not prove that it always wins. Nor does controlling total energy suffice to control what happens in an increasingly small region. This gap between global control and local behavior is central to Terence Tao’s explanation of the problem’s difficulty.
Clay’s statement has an important asymmetry
Charles Fefferman’s official statement offers four alternatives. They combine two types of result with two spatial settings:
| Result | Whole space | Periodic space | External force |
|---|---|---|---|
| Smooth solutions exist for all time | A | B | Set to zero |
| Breakdown of the required global existence | C | D | Permitted, under precise conditions |
Periodic space can be pictured as a box whose pattern repeats in every direction. It lets us study the dynamics without introducing walls.
The crucial detail is in the last column. The regularity route requires proving that every admissible initial condition produces a smooth solution forever without external forcing. The breakdown route allows a counterexample with external forcing.
These two routes are therefore not a statement and its exact negation. The negation of “all unforced flows remain smooth” would be to find an unforced flow that does not. Allowing an intervention enlarges the class of flows in which we can seek a counterexample.
A proof with external forcing would not, by itself, refute universal regularity in the unforced case. Another argument would be needed to connect the two. That is the potential source of dissatisfaction: the prize statement and the question we ask about the fluid have different scopes.
Why introducing a force does not make the result trivial
The objection is reasonable: if we can intervene in the fluid, are we not artificially causing the breakdown?
We are choosing an intervention, but we cannot smuggle in the singularity. Applying a force that becomes infinite and then pointing out that the velocity also blows up would not count.
“Smooth” does not capture all of Clay’s conditions either. In whole space, the force and its derivatives must decay rapidly in space and time; the periodic setting requires spatial periodicity and temporal decay conditions. The force must satisfy the assumptions even across the proposed breakdown time.
The main theorem in OpenAI’s manuscript makes a specific claim: for any positive viscosity, it constructs an evolution that starts from rest, keeps its kinetic energy bounded and develops unbounded velocities. The force is smooth and compactly supported in space and time: it acts only within a bounded region and time interval.
The introduction describes a vortex that concentrates and accelerates. The challenge is to balance the equation’s terms so that, although some grow without bound, their combination leaves a regular force. Oscillatory corrections in the construction aim to cancel the singular parts that would otherwise appear in that force.
This describes the submitted argument; it does not validate its steps. But it explains why allowing a force does not turn the problem into an easy exercise: the difficulty is keeping the force regular while the motion ceases to be so.
The question that remains when we write f = 0
Once we understand the intervention’s role, the next step seems obvious: remove it.
In the equation, that only requires writing f = 0. In the proof, it can change everything. The force is part of the construction; deleting it does not automatically preserve the behavior we obtained.
The question becomes whether the initial motion, evolving under Navier–Stokes without any external force, can generate a singularity by itself. A counterexample that needs forcing does not decide this.
We must also take the other possibility seriously: the unforced case might have a protection that does not survive when interventions are allowed. A forced breakdown result, considered on its own, does not rule that out.
None of this would make the advance small. Showing that a class of evolutions admits singularities would change our knowledge of the equations. But the next discovery would come from understanding which part of the mechanism depends on the force and whether it can be dispensed with, rather than simply repeating that the problem is solved.
A possible singularity does not describe every fluid
Proving that a case exists does not tell us how often it occurs. A construction may depend on extraordinarily precise conditions. Small perturbations might destroy its mechanism; a broad family of flows might preserve it. Existence alone does not tell us which is true.
Nor does it mean real water reaches infinite speed. Navier–Stokes is a continuum model. Turning a construction into a physical claim requires checking whether its assumptions remain appropriate at the scales involved.
Another useful distinction: losing smoothness does not mean losing every mathematical description. Weak solutions express the equations through integral relations and allow less regularity. Under appropriate conditions, the theory begun by Leray provides global weak solutions.
The breakdown of a smooth solution does not erase that theory. It opens questions about continuation, the energy conditions retained and the uniqueness that can be established in the chosen class of solutions. Different notions of weak solution provide different guarantees.
What the dispute with the mathematicians concerns
There is also a human history behind the announcement that should not be erased. In his public statement, Buckmaster describes his collaboration with Alpöge as personal, although Alpöge works at Anthropic, and credits Diego Córdoba and Luis Martínez-Zoroa with the original idea.
He alleges that Sébastien Bubeck asked for Alpöge to be excluded from an authorship proposal and recounts comments about the consequences for his career of making the dispute public. These are Buckmaster’s claims about private conversations. He also asks whether his Codex sessions could have fed into training, while explicitly saying he does not know whether his data was used.
OpenAI acknowledges that rumors about those advances prompted its initiative. It denies accessing the work before publication, while not ruling out that de-identified product usage data helped improve its models. It maintains that the proofs differ and recognizes the researchers’ priority on forced Euler. That is the company’s position.
These responses leave questions to clarify. Uncertainty about training does not prove appropriation of results; an account of the technical process does not, by itself, settle the authorship allegations either. The chronology, origin of ideas and terms proposed to the authors need their own evidence.
The tension extends beyond this case: a researcher may use tools from a company that also competes to produce advances in their field. A correct proof would not settle that question of trust. Questionable conduct, if established, would not by itself refute the mathematical argument either.
Solving should also help us understand
Advances do not come from nowhere. In his 7 September explanation, Tao places Alpöge and Buckmaster’s work within an earlier line of research by Córdoba and Martínez-Zoroa. The results he discusses concern other fluid equations: incompressible porous media, Boussinesq and Euler. Euler leaves out viscosity; progress there does not automatically solve Navier–Stokes.
Understanding what transfers between those equations is part of the proofs’ scientific value. So is determining the role of forces, which singularities are stable and what happens after smoothness is lost.
Acceptance of the result has its own process. Before considering a proposed solution, Clay’s rules require publication in a qualifying outlet, at least two years since publication and general acceptance by the mathematical community. That procedure is distinct from examining today whether the argument is correct.
If the proof is confirmed, we will have learned something profound: the conditions admitted in the construction do not suffice to protect the fluid from a singularity. The next question will be what produces that breakdown and how much it depends on external intervention.
Solving the problem establishes a conclusion. Understanding the proof should let us explain why it happens and how far it reaches. That is where work begins that no “Navier–Stokes solved” headline can declare finished.
Sources
- OpenAI: announcement and response on concurrent work, 8 September 2026.
- OpenAI: “Finite time blowup for Navier–Stokes” manuscript.
- OpenAI: Lean formalization repository.
- Charles Fefferman / Clay: official statement, including errata.
- Clay: Millennium Prize Problem rules.
- Tristan Buckmaster: statement on the results and conversations with OpenAI.
- Terence Tao: why global regularity for Navier–Stokes is hard.
- Terence Tao: smooth forcing advances in related equations, 7 September 2026.
