A Weekend, a Model, and the Navier–Stokes Millennium Prize: Inside AI's Messiest Math Breakthrough

For the second time in history, one of the seven Millennium Prize Problems — the million-dollar questions the Clay Mathematics Institute posed in 2000 — may have fallen. Unlike the first (Perelman’s proof of the Poincaré conjecture, which he declined to accept), this one was not written by a lone genius in seclusion. OpenAI announced on September 8, 2026 that an internal research model produced a proof that the Navier–Stokes equations — the 200-year-old partial differential equations that govern how fluids move — admit solutions that “blow up” in finite time. And it landed in the middle of one of the messiest priority disputes modern mathematics has seen, because the announcement came hours after an NYU mathematician published a detailed statement accusing the company of racing to preempt his own work.

What was announced

OpenAI’s post, titled “On the Navier–Stokes Millennium Prize Problem,” describes an “AI-generated solution” including both a human-readable writeup and a formal proof in Lean, the proof assistant language that machine-checks every logical step. The result targets a specific formulation: existence of forced blowup in R³ and T³ (three-dimensional Euclidean space and the three-torus), with the forcing function smooth under options (c) and (d) in Charles Fefferman’s official statement of the Clay problem.

The stakes are straightforward to state even if the mathematics is not. Fefferman’s statement asks for one of two things: either prove that smooth initial data always yields smooth solutions for all time, or exhibit initial data whose solution develops a singularity — a point where the velocity becomes infinite, something physically impossible. OpenAI claims the latter, for the forced version of the problem. If the proof survives scrutiny, it would be only the second Millennium Problem resolved in the 26-year history of the list.

Scientific American’s summary of the technical claim: on rare occasion, the equations “blow up,” meaning they dictate that a fluid’s speed becomes infinite at some points. The company says the proof has been certified in Lean, “which all but guarantees its correctness” — a stronger verifiability standard than any human referee process can offer, though formal verification confirms the chain of logic, not that the statement proven is the one the Clay Institute actually asked about.

The other half of the story

The announcement did not arrive in a vacuum. The night before, Tristan Buckmaster, a mathematician at NYU’s Courant Institute, posted three preprints with Levent Alpöge, a mathematician who works at Anthropic: finite-time blowup with smooth forcing for the incompressible porous medium equation, the two-dimensional Boussinesq system, and — most significantly — the three-dimensional incompressible Euler equations. The Euler result is widely viewed as a stepping stone toward Navier–Stokes: Euler is Navier–Stokes without viscosity, and a blowup for Euler with smooth forcing points directly at the harder problem.

Alongside the preprints, Buckmaster released a personal statement, and it is extraordinary reading. His account, in brief:

• September 3: With a rumor circulating that “Anthropic had resolved a major open problem,” Buckmaster emailed a prominent mathematician at OpenAI to clarify that the work was a personal collaboration with no institutional involvement, that he paid for his AI tools out of his own research funds, and that the papers would be “posted shortly, the paper and the formalization together.”
• September 6: After two scheduling attempts, he spoke twice by phone with that mathematician and Sébastien Bubeck, OpenAI’s VP of research. He was told an internal OpenAI model had produced a roughly 100-page proof of finite-time blowup for the forced Navier–Stokes equations. When Alpöge asked by text for the precise statement, the answer was: “Existence of forced blowup in R³ and T³,” with smooth forcing under Fefferman options (c) and (d).
• The “very little human input” framing unraveled over the call, Buckmaster writes. As team members fed corrections to Bubeck over internal chat, it emerged that an entire OpenAI team had worked the problem, had first set the model on easier cases including Euler, that the displayed prompt itself had been written by prompting Codex, and that “an insane amount of compute” had been used. The first prompt, he says it was eventually agreed, was sent “in the past few days” — after information about the pair’s work had reached OpenAI.
• Buckmaster asked whether the model had been trained on or had access to the pair’s Codex sessions, into which they had placed all their drafts for the entire project. He was told the model does not look up user data; on training, “I did not get an answer.”
• Two proposals were put to him, per the statement: either the pair posts Euler and OpenAI posts Navier–Stokes the next day, or Buckmaster alone writes the paper presenting the Navier–Stokes result while acknowledging an internal OpenAI model resolved it. Buckmaster says Bubeck twice pushed to have Alpöge removed from authorship, because Alpöge works at Anthropic — OpenAI’s chief rival in the frontier-model race.

When Buckmaster said he would go public if OpenAI released the result as proposed, the reply, as he recounts it, was: “Why would you ruin your career?” And after he asked why going public would ruin his career: “If you don’t want me to be nice, then I don’t have to be nice.”

Buckmaster is careful about what he is not claiming. He has not seen OpenAI’s proof. He does not know what their model did, or how. He does not know whether their data was used. “I am not accusing anyone of anything,” he writes. “I am stating what I was told, when, and what was proposed to me.”

OpenAI’s response

At a press briefing on Tuesday, Bubeck denied the rumors about the proof’s origins. His framing, per Scientific American: OpenAI’s internal model had independently solved the Euler problem “by totally different means” than Buckmaster and Alpöge’s approach, but the full Navier–Stokes proof did follow a similar method to theirs. And he confirmed the proof was developed over the weekend — that is, after the time Buckmaster says news of the pair’s result reached OpenAI. He categorically rejected any influence: “We did not use their prompt or proofs to prompt our models.”

Notably, the timeline itself is not really in dispute. The disagreement is over what it means. OpenAI says a model, pointed at a freshly important problem, closed it in days. Buckmaster’s concern is the direction: the route through smooth forcing was the Córdoba–Martínez-Zoroa program — a line of attack “almost nobody else” was working on, and the one he and Alpöge had quietly chosen. When he heard “forced,” he writes, “it was a bright red flag.”

Diego Córdoba, who with Luis Martínez-Zoroa developed the forcing approach the AI ultimately used, told Scientific American: “We’re a little bit in shock. If it’s done, that will be a big surprise for us.”

The uncomfortable middle: both things are true

The easy read is villains and victims, but the record supports something more complicated — and more interesting.

The mathematics is genuinely historic if it holds. A formalized proof of forced Navier–Stokes blowup, produced by a model over a weekend, would be the most consequential AI-assisted theorem ever. And Lean certification changes the epistemics: whatever one thinks of OpenAI’s conduct, the artifact itself is machine-checkable in a way Perelman’s arXiv preprints were not.

The conduct questions are genuinely serious. An unresolved question about whether the internal model was trained on users’ Codex sessions — where Buckmaster and Alpöge had stored every draft of their year-long project — now hangs over the most important math announcement of the decade. “I did not get an answer” is not a reassuring response to that question. The authorship proposals Buckmaster describes, if accurately recounted, would have rewritten the credit for a Millennium Prize result between two rival companies’ employees.

The precedent is the real problem. Buckmaster’s deeper point is about what the episode portends: “The significance of this with respect to the way we train students, assign credit, referee, and decide what is worth one human life’s attention cannot be understated. This is a Deep Blue–Kasparov moment.” A mathematician-plus-LLM collaboration can now do in a month what once took the field’s best years. When a lab hears a rumor of an outside breakthrough and can marshal “an insane amount of compute” at the same problem within days, independent verification of priority, provenance, and data boundaries stops being academic etiquette and becomes infrastructure the field lacks.

What happens next

The Clay Institute’s official problem page continues to list Navier–Stokes as unsolved, and prize consideration requires years of community scrutiny and published verification — the Poincaré conjecture took roughly four years from preprint to consensus. Reconciling two proofs whose lineage is “famously difficult to trace” (in Scientific American’s phrase) will take time. University of Chicago’s Luis Silvestre: “Yesterday and today are crazy days. We’re all, in the community, discussing the implications of this.”

The episode also lands amid a season of AI math milestones — Claude’s Riemann zeta bound at Anthropic in August, a Jacobian conjecture counterexample from Fable 5 in July, an Erdős problem in May — but none of those carried a contested origin story. This one does, and that may be its most lasting effect: the first time an AI theorem of this scale arrives with an asterisk attached not to its correctness, but to its provenance.

Sources are listed in the frontmatter of this post.