OpenAI’s Navier–Stokes claim is public, machine-checked and still awaiting independent judgment

• OpenAI’s paper constructs, for every positive viscosity, a smooth force and a flow that starts at rest, develops unbounded velocity before time 1 and retains bounded kinetic energy. ^[1]
• The accompanying Lean repository is public and includes build instructions. A successful formal check would verify the encoded theorem inside the Lean environment, not the historical provenance of the ideas or the broader research-integrity allegations. ^[2]
• Clay still lists Navier–Stokes as unsolved and requires a qualifying publication, two years of elapsed time and general community acceptance before considering a prize claim. ^[3]
• OpenAI says it did not access Buckmaster and Alpöge’s specific user data, while acknowledging that it cannot rule out de-identified data derived from product use helping improve its models. ^[4]
• Charles Fefferman, who wrote the official problem statement, told El País that the result appears to match the problem as formulated but still needs a conventional mathematical rewrite and detailed review by several experts. ^[5]

OpenAI has published a 166-page proof claiming that a smooth three-dimensional fluid can develop a finite-time singularity while retaining bounded kinetic energy, along with a Lean formalization intended to make the argument mechanically checkable. The company says the construction establishes alternatives C and D in the Clay Mathematics Institute’s official Navier–Stokes problem. [1]

That publication changes the status of the claim, but it does not make the announcement an accepted Millennium Prize solution. Clay’s own page continues to label the problem “Unsolved,” and its rules require a qualifying publication, at least two years of elapsed time and general acceptance by the mathematics community before the institute will consider a proposed solution. [3]

What OpenAI’s paper claims #

The manuscript does not claim that every unforced Navier–Stokes flow becomes singular. It constructs, for every positive viscosity, a smooth force and smooth initial data for which no global smooth solution with uniformly bounded kinetic energy exists. The construction starts with zero velocity, concentrates an inward-spiraling vortex near the origin and drives the velocity toward infinity as time approaches 1. [1]

The external force is central to the construction. OpenAI says singular pieces of the momentum balance cancel through oscillatory pulses and successive corrections, leaving a force that remains smooth even as the velocity becomes unbounded. The paper identifies the result with Clay’s alternatives C and D: breakdown on three-dimensional Euclidean space and on the periodic three-torus. [1][6]

This is a narrower and more precise claim than the headline “AI solved Navier–Stokes” suggests. It is a proposed counterexample to the global smoothness alternatives in the official formulation, not a proof that ordinary, unforced fluid flows generically blow up. Whether the construction satisfies every condition in that formulation is part of the mathematical review still to come.

What Lean verifies—and what it does not #

OpenAI has released a public repository containing Lean files for its Navier–Stokes and Euler results. The repository specifies Lean 4.34.0-rc2, Mathlib and Lake, and instructs users to retrieve the Mathlib cache and run the build. It also provides instructions for checking the formalizations with Comparator. [2]

A successful Lean build would be strong evidence that the formalized theorem follows from the definitions and axioms accepted by the Lean environment. It would not, by itself, establish that the formalized statement faithfully captures the intended Clay problem, that every translation from the paper is correct, or that the underlying ideas were independently derived.

That distinction matters in a formalization of this size. Reviewers must inspect the statement files, regularity and energy conditions, force construction and the link between the encoded theorem and alternatives C and D. The repository makes those materials available, but this desk has not independently compiled the project or completed a mathematical audit of the proof.

The official verdict is not in #

Charles Fefferman, who wrote the official Clay problem statement, gave the clearest public assessment located in this review. He told El País that OpenAI’s result matches the problem as he formulated it, while adding that he would need a rewrite in conventional mathematical prose and a close reading by several experts before becoming fully convinced. He estimated that process would take months. [5]

That is stronger than treating the paper as an unexamined model output, but it is not formal acceptance by Clay. The institute’s rules say it does not accept direct submissions and will consider a proposed solution only after publication in a qualifying outlet, two years since publication and general acceptance in the global mathematics community. OpenAI says it does not intend to claim the $1 million prize. [3][4]

The most accurate description, as of September 10, 2026, is therefore a public, machine-formalized proposed resolution whose correspondence to the Clay problem has received an encouraging preliminary assessment from its original author but not independent community acceptance.

The provenance dispute remains separate #

OpenAI says the effort began on September 1, 2026, after researchers heard rumors that two Millennium problems had been solved. The company says the Navier–Stokes effort involved about 10,000 concurrent agents, reached a resolution on September 5 after roughly 88 hours, and required another 17 hours for Lean formalization and verification. It reports 2.7 million messages and about 130 billion output tokens for the Navier–Stokes work. [4]

The timing became contentious because Tristan Buckmaster, a New York University mathematician, and Levent Alpöge, an Anthropic researcher, were working on related fluid-dynamics problems. WIRED reported Buckmaster’s allegation that OpenAI learned of their private work, accelerated its own effort and proposed a publication arrangement that would have excluded Alpöge from authorship. OpenAI executives denied using their prompts or proof to direct the system. [7]

OpenAI’s written account makes a narrower concession: it says no specific user data was accessed, but that the company cannot rule out de-identified data derived from product use helping improve its models. That does not establish that Buckmaster and Alpöge’s unpublished work influenced the proof, nor does it eliminate the question. Resolving it would require records of model training, product-data access, retention and the chronology of the researchers’ interactions—evidence not publicly available in the sources reviewed here. [4][7]

The credit dispute also reaches beyond Buckmaster and Alpöge. El País reported that OpenAI’s first version of its announcement omitted references to Diego Córdoba and Luis Martínez-Zoroa, whose prior work supplied a strategy involving cascading vortex layers; the newspaper said those references appeared several hours later in an updated version. Fefferman told the paper that the underlying ideas belong to a broader chain of work, while characterizing the immediate dispute as one between OpenAI and the Buckmaster–Alpöge team. [5][8]

Formal verification can help answer whether a stated theorem follows from its encoded premises. It cannot determine who first developed an approach, whether unpublished work influenced a model indirectly or whether the credit offered to collaborators was appropriate. Those are provenance and research-integrity questions, and they remain unresolved.

Companies mentioned #

Further sources #

The stories that matter, in one email. Free — unsubscribe anytime.