On September 8, OpenAI published a claimed solution to one of the seven Clay Mathematics Institute Millennium Prize Problems: Navier-Stokes existence and smoothness, the question of whether smooth, finite-energy fluid flow can spontaneously blow up in finite time. An internal, unreleased model ran roughly 10,000 agents concurrently for 88 hours to reach a proposed resolution, then spent another 17 hours formalizing the result in the Lean proof language. That’s a genuinely large claim, on a problem that’s sat unsolved since the Institute put a $1 million bounty on it in 2000.

Here’s the part that didn’t make it into the celebratory headlines: the Clay Institute hasn’t accepted it, and it’s still listed as open. And the origin story OpenAI told for why it rushed the effort in the first place has turned into a credit fight ugly enough that Axios wrote it up as its own story.

This is the second major AI-proof announcement in under a week. We covered Claude formalizing Fermat’s Last Theorem on September 5. Three days before that, OpenAI told the world to get ready for the “AGI era” at the GPT-6 Astra launch. Now there’s a third data point, and it complicates the story more than it confirms it.

Quick Summary: What Happened

Detail	Info
Date	September 8, 2026
The claim	An internal OpenAI model found a proof that smooth 3D fluid flow can develop a finite-time singularity under a version of the Navier-Stokes equations
Compute	~10,000 agents running concurrently for 88 hours, plus 17 more hours of Lean formalization
Clay Institute status	Has not accepted the result; Navier-Stokes remains listed as unsolved on claymath.org
The catch	The proof relies on a smooth external force — a version of the problem the Clay statement permits, but most mathematicians treat the unforced case as the real question
The dispute	OpenAI says it started the project after hearing a rumor about NYU’s Tristan Buckmaster and Anthropic’s Levent Alpöge — whose actual result addressed a different, related problem
Official source	OpenAI: On the Navier-Stokes Millennium Prize Problem

Bottom line: OpenAI has a real, Lean-checked result on a real version of a famous problem. It is not the version the Clay Institute is going to hand over a check for, and the way OpenAI got there is now a bigger story than the proof itself.

What Actually Happened

The Navier-Stokes equations describe how fluids move — water in a pipe, air over a wing, weather itself. They’ve sat in textbooks since the 1800s, and nobody has ever proven whether a smooth, well-behaved starting flow always stays smooth, or whether it can spiral into a singularity — a point where velocity or pressure shoots to infinity in finite time. That’s the Clay Institute’s Navier-Stokes existence and smoothness problem, one of seven Millennium Prize Problems named in 2000, six of which remain open.

Per OpenAI’s own writeup, the effort started September 1. An internal model — OpenAI describes it as meaningfully more capable than GPT-6 Astra, which had shipped publicly just two days earlier — coordinated close to 10,000 agents working in parallel. By Saturday, September 5, about 88 hours in, the agents converged on a proof that an initially smooth, finite-energy flow with a smooth applied force develops a singularity in finite time. Another 17 hours went into translating that result into Lean, where it compiled and checked out. OpenAI says GPT-6 Astra handled the formalization pass.

Worth sitting with that number for a second: 10,000 concurrent agents is an order of magnitude beyond the “several dozen” Claude used to formalize Fermat’s Last Theorem three days earlier. Fortune reported compute-cost estimates ranging from $2 million to $22.5 million for the run — an order of magnitude or more above what prior AI math efforts have cost. This wasn’t a clever prompt. It was industrial-scale search.

What Are the Four Versions of the Navier-Stokes Millennium Prize Problem?

The Clay Institute’s official problem statement isn’t one question — it’s four, and which one you solve changes everything about whether you’ve “solved Navier-Stokes” in the way the prize money implies:

1. Option A — Existence and smoothness of solutions in all of 3D space (R³), with no external force.
2. Option B — Existence and smoothness of solutions on the 3D torus (a periodic domain), with no external force.
3. Option C — The same question in R³, but allowing a smooth external force applied to the fluid.
4. Option D — The same question on the torus, with a smooth external force.

OpenAI’s proof addresses the forced case — Options C and D. That’s a legitimate part of the official problem statement, not a loophole OpenAI invented. But it’s also the version most working mathematicians in the field consider secondary. The unforced case, where the fluid moves purely on its own momentum with nothing pushing it, is the one that’s driven a century of research and the one nobody expects to fall soon.

Why the Clay Institute Hasn’t Signed Off

Clay Mathematics Institute president Martin Bridson called the announcement “exciting” but said the Institute’s evaluation would be “deliberately unhurried” and “absolutely rigorous.” Translated out of institute-speak: nobody’s cutting a $1 million check off a press release, no matter how many agents wrote it or how clean the Lean compile looks.

That’s not institutional foot-dragging. A Lean proof checks that a formal argument follows from its stated axioms — it doesn’t check that the formalization actually captures the problem as posed, and it says nothing about whether the argument extends to the unforced case anyone actually wants solved. Independent mathematicians haven’t finished walking through OpenAI’s roughly 100-page argument yet. Fields Medalist Terence Tao offered a sharper frame to CNN: reading a machine-generated resolution like this is “a little like going to watch a movie and jumping straight from the first ten minutes to the last ten minutes; technically, all the plot lines are resolved, but most of the value of the experience was lost.”

Tao’s broader worry, laid out on his own site, is less about this specific proof and more about the pattern: labs pointing enormous compute at open problems and extracting answers without the intermediate insight that usually teaches the field something. He called it “strip-mining” open problems — technically effective, but liable to hollow out the process that trains the next generation of mathematicians.

The Credit Dispute That’s Now the Bigger Story

Here’s where it gets uncomfortable for OpenAI. The company says it launched this project on September 1 after hearing a rumor about work by NYU mathematician Tristan Buckmaster and Anthropic researcher Levent Alpöge. The problem: what Buckmaster and Alpöge had actually proven, on August 15, was finite-time blowup with smooth forcing for the 3D Euler equations — the frictionless cousin of Navier-Stokes, not Navier-Stokes itself. Related. Not the same problem, and not the one OpenAI ended up claiming credit for solving first.

According to Axios’s reporting, OpenAI’s Sébastien Bubeck called Buckmaster on September 6 and described an internal, roughly 100-page proof for forced Navier-Stokes — a narrow approach Buckmaster says almost nobody else in the field was pursuing, which is part of why he suspects OpenAI’s agents had visibility into his work. He asked directly whether OpenAI’s model had been trained on, or had access to, his private Codex sessions. He says he didn’t get a straight answer.

The proposed resolution, per Buckmaster’s account, was worse than the ambiguity: Bubeck allegedly offered him a choice, either publish jointly with OpenAI releasing its own paper the next day, or publish solo and claim any prize — but only if he removed Alpöge’s name from the paper, because OpenAI objected to Alpöge’s employment at Anthropic. When Buckmaster pushed back, he says he was told something to the effect of “why would you ruin your career?”

OpenAI’s public response walks a line between disputing the specifics and acknowledging the credit is owed. Bubeck has said flatly: “We did not use their prompts or proofs to prompt our models or direct our agents,” and OpenAI denies inspecting any private user data, while conceding that de-identified platform usage could plausibly have factored into training generally. Separately, Bubeck later stated: “I want to be extremely clear that we recognize the priority of Levent Alpöge and Tristan Buckmaster’s work.” A real concession — but one that arrived after the reporting, not before it.

Why This Matters

Two separate claims are getting flattened into one headline. Claim one: OpenAI’s agents produced a genuine, Lean-verified result on a real (if secondary) version of a named Millennium Prize Problem, at a scale of parallel agent coordination — 10,000 concurrent processes — that’s a real engineering milestone regardless of what the Clay Institute eventually decides. Claim two: this proves OpenAI has leapfrogged into a new tier of mathematical capability that validates the “AGI era” framing from the GPT-6 Astra launch five days earlier. The first claim is defensible. The second is doing unearned work, and the credit dispute is exactly the kind of detail that framing was designed to paper over.

We already saw this shape once this week — Astra’s own 99.9% benchmark score dropped to 62.7% under a neutral test harness within a day of launch. Extraordinary number first, methodology and attribution questions surface within days, more modest reality settles in by the weekend.

What Are Your Options Now

If you’re tracking AI math-proving progress for your own field — formal verification, financial modeling, anything with a Lean- or Coq-style proof assistant in the loop — the actual signal here isn’t the Navier-Stokes headline. It’s that 10,000-agent concurrent coordination on a hard formal problem now works well enough to produce a compiling proof in under four days. That’s the infrastructure story, and it’s real independent of what the Clay Institute ultimately rules.

If you’re evaluating vendor claims about “solving” open problems, ask which specific variant got solved before taking the headline number at face value. The Clay Institute’s own four-way split on Navier-Stokes is exactly the kind of fine print that separates a real result from a marketing framing.

If you’re a researcher worried about AI labs racing ahead of your own unpublished work, this dispute is a preview of a real operational risk. Buckmaster’s account — called days before a public launch and offered a credit-sharing deal contingent on a colleague’s employer — is worth knowing labs are capable of before you put unpublished results anywhere near their tooling.

The Bigger Picture

Three major AI-and-math or AI-and-reasoning claims landed inside eight days this month. GPT-6 Astra’s “AGI era” launch on September 3. Claude’s Fermat’s Last Theorem formalization on September 4. Now OpenAI’s Navier-Stokes claim on September 8. Every one of them followed a similar arc this site has now documented three times running: an extraordinary number lands first, a more careful read follows within a day or two, and the real result — while still notable — ends up narrower than the launch framing suggested.

The scale jump matters on its own terms, separate from the Clay Institute question. Anthropic’s Prove2Me system, which coordinated Claude’s FLT formalization, ran several dozen agents against a shared theorem graph. OpenAI just ran roughly 10,000 concurrent agents against an open problem and got a compiling Lean proof out the other end within days. Coordinating that many parallel processes toward one hard formal target without the effort collapsing into noise is an infrastructure result that outlasts this specific controversy — the same underlying trend behind Anthropic’s multi-agent coding harness, just applied at a scale nobody’s shown publicly before.

Our Take

We think OpenAI has a real result and a self-inflicted credibility problem sitting right next to each other, and the company’s own materials make both halves checkable. The Lean formalization is public, the forced-versus-unforced distinction is right there in the Clay Institute’s own problem statement, and Bubeck’s later statement recognizing Buckmaster and Alpöge’s priority is itself an admission the initial framing needed correcting.

What we’d push back on hardest is the timing. OpenAI didn’t stumble into a credit dispute by bad luck — it started this project over a rumor about someone else’s unpublished work, reached its own result inside a week, then approached that researcher with a proposal that would have required erasing his co-author over which company employed him. You don’t need to resolve who touched what data to find that sequence of events bad.

For readers deciding how much weight to put on the “AGI era” story: treat the Navier-Stokes claim as evidence that large-scale agent coordination on hard formal problems is a real, fast-improving capability. Don’t treat it as evidence OpenAI has claimed a Millennium Prize, because it hasn’t — the Institute said so itself, in writing, the same week.

Frequently Asked Questions

Did OpenAI actually solve the Navier-Stokes Millennium Prize Problem?

Not according to the Clay Mathematics Institute, which still lists it as unsolved. OpenAI’s internal model produced a Lean-verified proof for the forced version of the problem — one of four variants in the Institute’s official statement. Most mathematicians consider the unforced case the central open question, and OpenAI’s result doesn’t address it.

How did OpenAI’s AI produce the proof?

Per OpenAI’s announcement, an internal, unreleased model coordinated roughly 10,000 agents working concurrently for about 88 hours to reach a proposed resolution, then spent another 17 hours formalizing the argument in Lean, reportedly using GPT-6 Astra for that formalization pass.

What is the credit dispute about?

OpenAI says it began the project on September 1 after hearing a rumor tied to NYU mathematician Tristan Buckmaster and Anthropic researcher Levent Alpöge. That pair had actually proven a related but distinct result — finite-time blowup for the forced 3D Euler equations — on August 15. Per Axios, Sébastien Bubeck later offered Buckmaster a publishing arrangement contingent on removing Alpöge’s name, citing his Anthropic employment — which Buckmaster has publicly disputed.

What’s the difference between the forced and unforced Navier-Stokes problem?

The unforced case asks whether a smooth fluid flow, left alone with no outside force, can develop a singularity purely from its own internal dynamics — the version mathematicians have worked on for decades. The forced case allows a smooth external force pushing on the fluid throughout, which makes constructing a blowup example easier. Both are part of the Clay Institute’s official four-part statement, but the unforced case is the one most closely associated with “solving Navier-Stokes.”

How does this compare to Claude formalizing Fermat’s Last Theorem?

Both are Lean-verified formalizations produced largely by autonomous AI agents within days rather than years. The scale differs by an order of magnitude — Claude’s FLT effort used several dozen coordinated agents over 11 days; OpenAI’s Navier-Stokes run used roughly 10,000 concurrent agents over about four days total. Claude’s result formalized a known, previously-proven theorem; OpenAI’s addresses an open problem, but only in its forced variant, which is the source of the current dispute.

Has the proof been independently verified?

The Lean formalization compiles against Lean’s own kernel, confirming the logical steps follow from the stated axioms. That’s different from independent mathematical review of whether the problem was framed and translated correctly, which the Clay Institute has said is still underway.

Is the model that produced this proof publicly available?

No. OpenAI describes it as an internal, unreleased model more capable than GPT-6 Astra, which shipped publicly on September 3. There’s no public release timeline for it.

---

Last updated: September 10, 2026. Sources: OpenAI — On the Navier-Stokes Millennium Prize Problem · Clay Mathematics Institute — Navier-Stokes Equation · Axios · Fortune · CNN Business · Quanta Magazine.

Related reading: Claude Formalizes Fermat’s Last Theorem in 11 Days · GPT-6 Astra Lands: Inside OpenAI’s ‘AGI Era’ Claim · Anthropic’s Multi-Agent System: 4-Hour AI Dev · OpenAI’s Astra Crosses AI’s First Critical Cyber Line