OpenAI's Navier-Stokes Math Claim Ignites a Credit Fight

By Yash KharePublished Sep 9, 2026Updated Oct 30, 20265 min read

In the second week of September 2026, OpenAI said an unreleased internal model had produced a proof of finite-time blowup for the 3D Navier-Stokes equations, the fluid-dynamics problem behind one of the Clay Institute's $1 million Millennium Prize questions. Within hours, NYU mathematician Tristan Buckmaster said he and Anthropic researcher Levent Alpöge had been working on closely related results for months, and accused OpenAI of trying to push Alpöge off the credit. Updated October 30, 2026: here is what each side claims, what is actually verified, and how to judge it.

⚡ Quick facts

OpenAI's claim: an unreleased model, run as about 10,000 agents for about 88 hours, produced a forced-blowup proof for 3D Navier-Stokes, per OpenAI
The rival work: Buckmaster and Alpöge posted Lean-checked blowup proofs for three simpler fluid equations, including 3D Euler, in early September 2026
The dispute: Buckmaster says OpenAI's Sébastien Bubeck pushed to drop Alpöge because he works at Anthropic; Bubeck denies it but apologised for a remark
Clay status: the Clay Mathematics Institute still lists Navier-Stokes as unsolved and says its review will be unhurried
Key caveat: both results rely on a smooth external force, which the problem allows but which drives the method

What each team says it proved

Buckmaster (NYU) and Alpöge (an Anthropic researcher, working on this as a personal project) posted proofs of finite-time blowup with smooth forcing for three equations: incompressible porous media, the 2D Boussinesq system and the 3D incompressible Euler equations. "Blowup" means a smooth solution develops a singularity in finite time. Their proofs came with formal verification in Lean, and they say they used AI tools, including Claude and OpenAI's Codex, throughout.

OpenAI's claim is bigger: a forced-blowup result for the full 3D Navier-Stokes equations, which add viscosity. Per coverage in TechRepublic and The Batch, OpenAI says an internal model running about 10,000 agents worked for roughly 88 hours with little human input, and that the proof was also checked in Lean. Reported token and compute figures differ between outlets, so we do not repeat them.

The credit fight

Buckmaster says Bubeck twice pushed for Alpöge to be removed from authorship of a joint write-up because he works at a rival lab, and offered arrangements he describes as a bribe. Bubeck, in his own account reported by The Next Web, denies trying to strip Alpöge of credit for his work, says the proofs differ significantly, and apologised for a comment about Alpöge's career. Sam Altman backed Bubeck. OpenAI says its researchers and agents did not see the pair's work before it was public, though its write-up says it cannot fully rule out that de-identified usage data helped its models. OpenAI called Buckmaster's allegations false.

What is verified, and what is not

The Clay Institute has not accepted any proof. Its president said the evaluation would be "deliberately unhurried," per Implicator. A Lean check shows the proof is logically valid for the statement as encoded; experts still have to confirm that the Lean statement matches the Clay problem. The other open point is the external force: the official problem allows a forcing term, but critics note there is no evidence the construction works without one, which is the version most mathematicians care about.

Our take: a checklist for "AI solved X" claims

Is the statement the right one? A Lean proof is only as good as the theorem it encodes. Wait for experts to compare it with the official problem.
Is the proof public? Reports conflicted on how much of OpenAI's material was released. Only public, checkable proofs count.
Who did what? Both teams used AI heavily. Credit should follow the ideas and the verified work, not the employer.
Does the prize body agree? Until Clay rules, "solved a Millennium Prize problem" is a claim, not a fact.
Does the dispute change the maths? No. A correct proof is correct whoever wrote it; the fight is about provenance and fairness.

For a verification-first AI maths result, see our explainer on Claude's Fermat's Last Theorem formalization.

Sources

Frequently asked questions

Did OpenAI solve the Navier-Stokes Millennium Prize problem?

Not officially. OpenAI claims a forced-blowup proof for 3D Navier-Stokes, but the Clay Mathematics Institute still lists the problem as unsolved and says its review will be slow and rigorous.

Who are Buckmaster and Alpöge?

Tristan Buckmaster is an NYU mathematician and Levent Alpöge is a researcher at Anthropic. They worked together as a personal project and posted Lean-verified blowup proofs for three simpler fluid equations.

What is the credit dispute about?

Buckmaster says OpenAI's Sébastien Bubeck tried to have Alpöge dropped from authorship because he works at Anthropic. Bubeck denies that but apologised for a remark about Alpöge's career.

What does finite-time blowup mean?

It means a solution that starts smooth develops a singularity, such as infinite velocity at a point, after a finite amount of time.

Does a Lean proof settle the question?

It shows the proof is valid for the statement as written in Lean. Experts still need to check that the Lean statement matches the official problem.

Why does the external force matter?

The official problem allows a smooth external force, and both results use one. Critics note there is no evidence the methods work without it, which many mathematicians see as the harder question.