OpenAI Claims a Navier–Stokes Proof—Independent Review Is the Real Test

On September 8, 2026, OpenAI published a claimed solution to the three-dimensional Navier–Stokes existence and smoothness problem. In OpenAI’s technical account, the company attributes the proof to an internal AI system and presents an analytical manuscript plus a Lean formalization for a finite-time singularity.

Nature’s September 8 coverage treated the result as a company claim rather than an adjudicated solution. The immediate status is therefore clear: a specific proof artifact is available for scrutiny, but broad independent verification, qualifying publication and recognition by the Clay Mathematics Institute have not been completed.

What the proposed proof establishes

The Navier–Stokes equations describe fluid motion, and the Millennium Problem asks whether smooth solutions of the three-dimensional incompressible equations must remain smooth. A valid counterexample would instead show that permitted initial conditions and forces can drive a solution into a singularity, where velocity becomes unbounded within finite time.

The proposed construction begins with a smooth fluid at rest and applies a smooth external force. It produces a vortex that spirals inward and stretches along its axis as its central region contracts and its velocity grows without bound, while total energy remains finite.

The claim concerns alternatives C and D in the official problem formulation, which allow a smooth external force. That scope matters: the result is not a general assertion that ordinary water or air will reach infinite speed, nor is it a proof that every Navier–Stokes flow breaks down. It is a proposed mathematical counterexample within the specified continuum model.

If every step is correct and the construction satisfies all hypotheses in the official formulation, one counterexample is enough to resolve the problem in the negative. The unresolved issue is not what the manuscript intends to prove, but whether the argument actually meets that standard.

What Lean verification does—and does not—establish

Formalization is stronger than attaching executable calculations to a conventional paper. The Lean language reference explains that tactics construct proof terms, which are checked by Lean’s kernel and can also be examined by independently implemented external checkers.

A successful check establishes that the encoded conclusion follows from the encoded assumptions, definitions and dependencies. It can expose missing symbolic steps and prevent many forms of informal hand-waving, making the formal artifact an important layer of evidence.

But the kernel does not decide whether the encoded theorem is an exact translation of the Millennium Problem. Specialists must still inspect whether the formal definitions capture the required notions of smoothness, forcing, finite energy and singularity; whether imported results are appropriate; and whether the analytical manuscript and formal theorem make the same claim.

This creates two separate verification tasks. Independent teams need to reproduce the formal build and audit its dependency chain, then compare the machine-checked statement with the conventional mathematical formulation. Passing the first task cannot compensate for a scope or translation error in the second.

The verification ladder extends beyond formalization

A public manuscript begins review; it does not finish it. The relevant stages are distinct:

1. Public production: an analytical manuscript and formal proof files are available for examination.
2. Machine checking: the encoded derivation can be built and checked in its stated Lean environment.
3. Independent reproduction: external specialists rerun the formal proof, inspect dependencies and test the match between code and manuscript.
4. Expert mathematical review: analysts examine the construction, estimates, hypotheses and claimed singular behavior using ordinary mathematical scrutiny.
5. Qualifying publication: the proposed solution appears in an eligible refereed outlet.
6. Institutional consideration: sufficient time and community acceptance permit Clay to evaluate the result.

The final stages have explicit procedural constraints. The Clay Mathematics Institute’s rules require publication in a qualifying outlet, a wait of at least two years and general acceptance in the global mathematics community before the institute will consider a proposed solution; Clay does not accept direct submissions.

Those requirements make an immediate declaration of a settled Millennium Problem impossible, regardless of confidence in the formal files. Machine verification can accelerate scrutiny, but it does not replace journal review, sustained expert examination or the institutional timetable.

The authorship dispute is a parallel confidence test

The release is also shadowed by questions about concurrent unpublished work by New York University mathematician Tristan Buckmaster and Anthropic researcher Levent Alpöge. Axios’s account of Buckmaster’s objections documented concerns about whether knowledge of their research direction accelerated OpenAI’s effort, whether private Codex material could have influenced it and how authorship was discussed.

The projects should not be treated as identical. Buckmaster and Alpöge were pursuing a related forced Euler result, whereas the OpenAI materials distinguish an unforced Euler construction from the later forced Navier–Stokes claim. Concurrent timing and related techniques can justify a provenance inquiry without establishing that unpublished material entered the released proof.

On September 10, the public account was amended with an internal-investigation finding that Buckmaster’s Codex prompts from the preceding two months could not have influenced the system, including through training, and that the outside work was not viewed before publication. This is the organization’s internal conclusion, not an independent audit of model development, data access, project records or the chronology of ideas.

Provenance and validity answer different questions. A mathematically correct proof would remain correct despite an attribution dispute, but unresolved provenance can affect credit, trust in the development history and researchers’ confidence when placing unpublished work in commercial AI tools. Conversely, evidence of clean provenance would not validate the theorem.

Where the claim stands now

Produced and formalized: yes. An analytical proof and Lean artifact are public. Broadly independently accepted: not yet established. Published through the route required for Clay consideration: not established. Recognized as a completed Millennium Problem: no.

The next meaningful evidence will come from independent reconstruction of the Lean proof, expert comparison with the official problem statement, conventional mathematical review and any qualifying publication. The provenance controversy requires its own documentary examination. Until those tracks advance, the defensible description is a public, machine-formalized proof claim—not a settled Millennium Problem.

Also read:

• OpenAI Claps Back Hard at Anthropic with GPT-5.4-Cyber: Real Cyber Defense for the Actual World, Not Just the Elite
• Brand Storytelling in 2026: Authenticity Is Expected, and Claims Need Proof
• QuantHealth Raises $45M—Virtual Trials Still Depend on Real-World Proof