Solved, As Written: OpenAI's Navier-Stokes Claim and the Priority Dispute

On September 8, OpenAI announced that an internal model — run as roughly 10,000 concurrent agents for 88 hours, generating 2.7 million messages and about 130 billion output tokens, then 17 more hours formalizing in Lean with GPT-6 Astra’s help — had produced a proof of finite-time singularity formation for the Navier-Stokes equations, released as a 166-page paper and a public Lean formalization.¹ Within hours, NYU’s Tristan Buckmaster raised the obvious objection: he and Levent Alpöge — an Anthropic researcher working in a personal capacity — had spent a year on the neighboring problem, reached finite-time blow-up results for the Euler, Boussinesq, and porous-media equations under smooth forcing on August 15 with their own Lean verification, made the work public days before OpenAI’s announcement, and earned Terence Tao’s praise.² Buckmaster alleges the rumor of their unpublished progress is what kicked off OpenAI’s sprint; he kept his drafts in OpenAI’s own Codex, asked whether they had been seen, and describes the answers as evasive, then openly hostile.³ OpenAI denies direct access, concedes it cannot rule out anonymized user data in training, and its internal investigation cleared itself; it offered Buckmaster coauthorship on the press release — with Alpöge, the Anthropic-affiliated half of the pair, omitted.⁴ The Clay Mathematics Institute has not accepted the result and OpenAI says it won’t claim the prize.⁵ Four frames on what just happened.

First Principles: What Is a Proof?

Strip the announcement to fundamentals. A proof has two halves: a formal object that a machine can check, and a social process — who saw what, when, through whose instruments — that it cannot. Lean settles the first half and gives the claim a real chance of survival. The second half is where the story actually lives, and it is not formalizable. Scope matters too: the paper targets Clay options C and D, the forced variants; the unforced questions A and B that most mathematicians consider the real problem are untouched.⁵ In the field’s own words, “the Clay problem, as written, is solved. But the Clay problem, as many experts imagine it, lacks the piece that the forcing method relies on.”⁶ Rebuilt from fundamentals, the news is not “AI solved the Millennium Problem.” It is: an unverified manuscript exists, for a variant the field treats as a side door, authored by a party whose instruments the rival drafted in. The mechanical half of mathematics was never in dispute. The human half is the whole story.

Counterfactual: Did the Agents Matter?

Minimal intervention: delete OpenAI’s 10,000-agent sprint; hold everything else fixed. The forcing technique predates it — built by Diego Córdoba and Luis Martínez-Zoroa, then extended by Buckmaster and Alpöge, who paid for Claude and Codex out of research funds and got there first on the cousin problems.⁶ First-order consequence: forced Navier-Stokes blow-up still arrives, later — probable, not certain. Second-order: the lab-compute arms race compresses the gap regardless. Equilibrium check: compute scales the search, it does not originate the idea; the idea was already in the air. Verdict: the theorem was overdetermined; the timing was contingent on compute. Which is exactly why the announcement reads as it does — “in 88 hours” is the headline, not “blow-up.” What was new was not the mathematics but the time-to-mathematics, and time-to-X is a product claim wearing a lab coat.

A Nietzsche Ladder on the Priority Dispute

1. Camel

Mathematics carries an inherited code: priority is settled by timestamps, correspondence, and community judgment. Its instruments — paper, mail, the seminar — were neutral, owned by no claimant. The code exists because the founding dispute went so badly: Newton and Leibniz haunts every priority fight, and the lesson drawn was that neutral instruments and public verification keep calculus from devouring its children.

2. Lion (responding to Camel)

You carried the code well — now ask who owns the instruments. Buckmaster drafted inside Codex, a tool whose owner is the rival claimant, and asked the one question the code requires: did you see my work? The answers were evasive, then hostile; “cannot rule out” anonymized training data; the accused ran the investigation and cleared itself.³ The press release is the journal, the journal is the competitor, and the referee, the accused, and the beneficiary are one company. The code presumes symmetric actors with shared instruments. The symmetry is gone. No.

3. Child (responding to Lion)

Your No clears space for a new instrument. What the era needs is a lab notebook for models: provenance logs recording what a system saw and when, hash-stamped and auditable by someone who isn’t the claimant. Priority claims filed to neutral timestamps; laboratories of record outside the lab that benefits. This is not nostalgia for the seminar — it is the Camel’s old code, rebuilt for tools that remember everything and volunteer nothing. The Child’s game: make verification cheap again, so that proof alone can earn belief.

Inversion: How to Guarantee AI-Assisted Math Fails

The goal is AI as a trusted engine of discovery. How would you guarantee the opposite? Announce scope by headline — “Millennium problem solved” — while the paper targets the forced variant, and let C/D blur into A/B. Publish by press release before any independent verification, which Clay’s own process — two years and broad acceptance — exists to require.⁷ Store rivals’ drafts in your own instruments and stonewall when asked. Offer credit that splits the accusers: coauthorship for one mathematician, erasure for the other.⁴ Let the accused run the investigation. Five for five. Stated forward: precise scope claims, community verification before the verb “solved,” provenance logs, neutral priority mechanisms, independent review. None of it is expensive. All of it was skipped.

What the Frames Agree On

The theorem may survive review — Lean gives it that chance. But the frames converge on what the announcement was: maximum narrative extracted from minimum verified scope. Mathematics was the last discipline where verification was supposed to be cheap and trust in the producer unnecessary — where proof alone earned belief, no credentials required. AI-era mathematics quietly reintroduces trust in the producer exactly where it was never supposed to live, which is the Guardian’s verdict rendered institutional: humans are still vital, and the firms decline to see it.⁸ The real Millennium question is not whether machines can prove. It is whether institutions can audit — before the next 88-hour headline.

1. https://www.hpcwire.com/bigdatawire/2026/09/15/unsolved-for-90-years-openai-says-ai-cracked-millennium-prize-problem-in-88-hours ↩

2. https://www.implicator.ai/clay-institute-navier-stokes-openai-proof-claim ↩

3. https://techweez.com/2026/09/14/openai-navier-stokes-proof ↩ ↩²

4. https://the-tartan.org/2026/09/21/nyu-professors-and-openai-solve-navier-stokes-equations ↩ ↩²

5. https://www.datacamp.com/blog/openai-navier-stokes-math-problem ↩ ↩²

6. https://aiweekly.co/alerts/openai-mathematician-clash-over-ais-navier-stokes-proof-claim ↩ ↩²

7. https://modelcurrent.com/article/openai-navier-stokes-proof-review ↩

8. https://www.theguardian.com/commentisfree/2026/sep/20/the-guardian-view-on-ai-v-mathematicians-humans-are-still-vital-to-the-field-but-tech-firms-refuse-to-see-that ↩