An AI system says it broke a 90-year-old problem. We read the proof, and the small print is the story
Ten thousand agents, 88 hours, 166 pages and 641,000 lines of Lean. The mathematics is more serious than the sceptics allow, the credit is messier than the headlines allow, and nobody outside OpenAI has checked it yet.
On 8 September 2026 OpenAI published a 166-page proof that a three-dimensional fluid starting from rest can develop unbounded velocity in finite time, together with a Lean formalization of it. The work came from an unreleased internal model it describes as significantly more capable than GPT-6 Astra, running on the order of 10,000 concurrent agents, which reached the result on 5 September, 88 hours after launch.
The question is about as old as modern analysis gets. Jean Leray proved in 1934 that solutions exist in a generalized sense, and whether they always stay smooth became the open problem. Clay attached a million dollars to it in 2000, and Charles Fefferman wrote the official statement as four alternatives. OpenAI claims the negative ones, C and D.
We went and read the artifacts rather than the coverage: the theorem statement, the Clay problem description it points at, and the 641,000 lines of Lean. The result is stronger than the sceptics have been saying, weaker than the headlines, and the one thing nobody can yet claim is that it has been checked.
What the theorem says, what the official statement permits, and what is actually in the repository
1. The theorem, as written
Theorem 1.1 of the paper: for every viscosity above zero there is a smooth force compactly supported in space and in strictly positive time, and smooth velocity and pressure fields on the interval up to time 1, starting from zero initial velocity, everything supported inside one compact set, whose energy stays bounded for the whole interval while the peak speed runs to infinity as time approaches 1. Consequently no smooth global solution with bounded energy exists for that force and that initial state. The object is a vortex that spirals inward and stretches out like spaghetti, its core shrinking as it accelerates. The paper says this establishes alternative (C), and that compact support gives the periodic version (D) as a corollary.
2. The forcing objection fails, and it is worth seeing exactly why
The most repeated criticism is that the proof cannot count because the Clay problem forbids an external force. Read Fefferman's statement: f must be identically zero only in the positive alternatives (A) and (B). The negative ones do not require it. (C) asks for a smooth divergence-free initial field and a smooth f satisfying condition (5), which demands that f and all its derivatives decay faster than any power of (1 + |x| + t). OpenAI's force is smooth and compactly supported in space and time, which satisfies that condition outright: outside a bounded region the force is not small, it is zero. The same compact support gives the bounded-energy requirement. Fefferman explains the design himself: the four alternatives exist “to give reasonable leeway to solvers while retaining the heart of the problem.” Whatever is still open here, the presence of a force is not it.
3. Ninety years, or twenty-six, depending on which clock you read
OpenAI dates the question to Leray in 1934 and calls it roughly 90 years, which is the honest way to count the mathematics. The prize is much younger: Clay posed its version in 2000, making the Millennium problem twenty-six years old. The equations themselves are nineteenth-century, from Navier and Stokes. All three numbers circulated this week, mostly without anyone saying which was meant.
4. How the run worked, in OpenAI's own account
It had been training the internal model since 28 August. On 1 September it heard rumours that two Millennium problems had been resolved and launched agents at every open Millennium problem and a few others. Different groups got different variants of the statement, some aimed at proving (A) and (B), others at disproving through (C) and (D), with tools to run code and read a cached copy of the internet. Groups were cross-pollinated by using Codex to consolidate their intermediate insights, and were upgraded mid-run when a further-trained model arrived. The Navier-Stokes group ran about 10,000 agents, sent 2.7 million messages and burned roughly 130 billion output tokens; across all problems, 4.9 million messages and about 300 billion tokens. GPT-6 Astra then took 17 hours to produce and verify the Lean formalization.
5. The result that got a footnote has no force at all
Before Navier-Stokes the same system was set on what OpenAI calls easier problems, and about 100 agents working roughly 50 hours resolved the regularity question for the Euler equations, the same system with viscosity switched off. That paper is 57 pages and its result is unforced: smooth, compactly supported, divergence-free initial data whose solution blows up in finite time, with the velocity's C¹ norm unbounded and the time integral of the vorticity's supremum divergent. No force, no objection available. It is a famous open problem in its own right, it appears in the announcement as a stepping stone to the bigger one, and a year ago it would have been the story of the year.
6. The race was started by a rumour, and OpenAI says so in print
The page is unusually candid: the effort began on 1 September after hearing a rumour the company later realised concerned Levent Alpöge, an Anthropic employee, and Tristan Buckmaster, a mathematics professor at NYU. “Inspired by these rumors and by the step change in performance of our internal model, we launched an effort to evaluate it on all open Millennium Prize problems.” Whatever else is disputed, the trigger is not. The machine was aimed at this problem because humans were known to be close to it.
7. The mathematics did not start this month, and the paper knows it
Diego Córdoba and Luis Martínez-Zoroa built the programme underneath all of this, constructing singularities by amplification across scales while keeping the external force under control. Alpöge and Buckmaster spent about a year pushing it into the far harder smooth-forcing regime, and by 22 August 2026 had Lean-verified proofs for the incompressible porous medium equation, the two-dimensional Boussinesq system and the three-dimensional Euler equations. They posted late on 7 September, about twelve hours before OpenAI. Charles Fefferman, who wrote the Clay statement, told Quanta the heroes of the story are Córdoba and Martínez-Zoroa. Worth noting against the accusation that OpenAI failed to credit its sources: the paper's related-work section cites their constructions directly and describes how its own mechanism differs. The announcement page, which is what most people read, names neither.
8. Two accounts of the same week, and we cannot adjudicate between them
OpenAI says it contacted the pair on 6 September to offer a concurrent release and to recognise their priority in a joint announcement, offered them visibility into all its prompts and later the proof, states that neither researchers nor agents saw their work before it was public, and reports an investigation confirming that Buckmaster's Codex prompts in the preceding two months could not have influenced the system, including through training. A dated footnote records that this section was rewritten on 10 September with the investigation's findings, which is also the paper trail for the wording narrowing over several days. Buckmaster's account of the same conversations is that he was offered a choice between publishing with reduced credit and claiming the prize with Alpöge's name removed because Alpöge works at Anthropic, and that he was asked why he would ruin his career when he threatened to go public. The accounts are not compatible and nothing available to us settles it.
9. What is actually in the Lean repository, because we looked
openai/NavierStokesAndEuler, Apache-2.0, published the day of the announcement and last updated on 10 September: 2,659 Lean files, about 641,000 lines, built against Lean 4.34.0-rc2 and Mathlib. We scanned it. The development contains no sorry placeholders, no declared axioms and no native_decide escape hatch. The only five occurrences of sorry sit in a directory called ComparatorChallenges and are deliberate: they are the theorem statements with the proofs left out, for an independent checker to discharge.
10. The verification design is the most under-reported thing here
The obvious failure mode for a machine-produced proof is not a broken step, it is a formal statement that quietly says something weaker than the problem. OpenAI's answer is to target somebody else's statement: the ComparatorChallenges files are adapted from Google DeepMind's Formal Conjectures project, which had independently written the Millennium problem's alternatives (C) and (D) in Lean, complete with the decay conditions on the force. The repository ships instructions for checking the work with Comparator, a Lean-community tool that exports the proof and re-checks it with external kernels in a sandbox. So the question “did they prove the real thing” has a mechanical answer available to anyone who runs it. Nobody outside the company has published the result of doing so.
11. What our scan does not establish
Reading a repository is not checking a proof. We did not run lake build, which needs the full Mathlib toolchain and hours of compute, so we have not machine-checked anything, and a static scan cannot see an unsound custom tactic or a definition that drifts from the intended meaning. The absence of a sorry is necessary, not sufficient. The honest status is that the artifacts needed for independent verification are public and unusually complete, and that independent verification has not yet happened.
12. What the Clay Institute can do about it, and when
Its rules are unromantic. A solution must appear in a refereed mathematics journal of worldwide repute, then survive two years of general acceptance in the community before the Scientific Advisory Board decides whether to convene a committee at all. The Navier-Stokes page still carries the status Active, and Clay's president has called the announcement exciting while describing the review as deliberately unhurried. On that timetable the earliest anything could be awarded is late 2028, and only if the work is published, correct and accepted.
13. Declining the prize is data, not modesty
“We do not intend to claim the Millennium Prize for this result.” Read generously, deference to the mathematicians whose programme made it possible. Read plainly, a claim not filed cannot be rejected. Either way it is the most informative line on the page about the company's own confidence, and it sits strangely under headlines reporting that the problem has fallen.
14. Terence Tao's objection survives the proof being correct
Tao wrote about the Alpöge and Buckmaster results on 7 September, called them remarkable and thought the methods might reach Navier-Stokes. His concern is not validity but ecology: that the indiscriminate strip-mining of open problems for solutions may destroy the environment in which the next generation of techniques, problems and mathematicians would have developed. Answers arrive, insight does not, and the failed attempts that teach a field are never written down. That criticism is about how the result was obtained and published, and it stands whether or not the mathematics holds.
15. What it does to our AGI Clock: about a month
We moved the clock 30 days nearer, the same weight we gave GPT-6 Astra's Critical cyber rating. The capability is real: days of sustained autonomous work on a research problem, ending in an artifact a machine can check, is a different class of evidence from a benchmark score, and it reveals an internal model well ahead of anything on sale. Against that, nothing has been independently verified, the credit is contested, and the agents were pointed down a path human mathematicians had built and were days from finishing. A larger delta would price in a verification that has not happened.
16. What it does to our Immortality Index: nothing, and that is the point
The index moves only on biological or human evidence, so this is logged as context and moves the curve by zero, exactly as GPT-6 Astra was. Nobody lived a day longer for it. It is recorded because the capability demonstrated, holding a research problem for days without human direction and producing something checkable at the end, is the kind that could later be turned on protein design or trial analysis. When that day comes the index will need a reason measured in outcomes, not in tokens.
The extrapolation worth making is not that mathematics is solved. It is that the bottleneck moved. The limiting question used to be whether a machine could produce research-level mathematics at all, and that question now has an answer with 641,000 lines of Lean attached. The new limit is whether anyone can check the output fast enough, fairly enough, and with enough credit reaching the humans whose programme the machine followed. The checking is at least mechanised, and someone still has to run it. The crediting is not mechanised at all, and this week showed what that costs.
The last paragraph of OpenAI's own page is the one almost nobody quoted: the company says it is focused on understanding this model and using what it learns “to help us guide and pace how we pursue further advances in capability.” Four days later Dario Amodei published an essay asking the industry to pace the frontier, and Sam Altman agreed with it. Read in that order, this is not only a capability demonstration. It is the exhibit both sides of that argument are now pointing at. Our clock moved a month on it. It will not move a year until somebody outside OpenAI has run the checker.
Where we track this
- The AGI Clock →
A published baseline plus a signed day-delta per news item, every movement traceable to a sourced card. This result is logged there at minus 30 days, with the reasoning you have just read.
- The Immortality Index →
A life-expectancy curve that moves only on biological or human evidence. This result appears on its timeline as context and moves nothing, which is the honest answer for a mathematics proof.
- Three AI bosses called for a slowdown →
Published 15 September: what Amodei, Altman and Musk actually committed to four days after this announcement, and why no model has been delayed.
Questions people ask
- Did AI solve the Navier-Stokes Millennium Prize problem?
- OpenAI says an internal model produced a proof that solutions can develop a singularity in finite time, which resolves the negative branch of the problem, and published a 166-page paper with a Lean formalization on 8 September 2026. It has not been independently verified, the Clay Mathematics Institute still lists the problem as Active, and OpenAI says it does not intend to claim the prize. The accurate statement today is that a serious, unusually checkable claimed resolution exists and nobody outside the company has checked it.
- Does the external force in the proof disqualify it?
- No. Fefferman's official Clay statement requires a zero force only in the positive alternatives (A) and (B). The negative alternatives (C) and (D), which is what OpenAI claims, explicitly permit a smooth force whose derivatives all decay faster than any power of (1 + |x| + t). The force in Theorem 1.1 is smooth and compactly supported in space and time, so it is zero outside a bounded region and satisfies that condition outright. The open questions are about correctness, not about the presence of a force.
- Has the proof been verified?
- Not independently, as of 16 September 2026. The Lean development is public under Apache-2.0 at openai/NavierStokesAndEuler, contains no sorry placeholders and no declared axioms, and ships challenge files adapted from Google DeepMind's Formal Conjectures statement of the problem so that an outside party can check the proof against a statement OpenAI did not write, using the Comparator tool. Running that check is the missing step, and building the project requires the full Mathlib toolchain.
- What did Buckmaster and Alpöge prove, and how does it relate?
- Tristan Buckmaster of NYU and Levent Alpöge, a researcher at Anthropic, posted three finite-time blowup results with smooth forcing on 7 September 2026, covering the incompressible porous medium equation, the two-dimensional Boussinesq system and the three-dimensional Euler equations, Lean-verified on 22 August. They extended a programme built by Diego Córdoba and Luis Martínez-Zoroa. Their work does not resolve Navier-Stokes. OpenAI's agents separately resolved the unforced Euler question, and its page recognises the pair's priority on the forced Euler case.
- When could the $1 million prize actually be awarded?
- Not before late 2028 at the earliest. Clay's rules require publication in a refereed mathematics journal of worldwide repute, then two years of general acceptance in the mathematical community, before its Scientific Advisory Board decides whether to consider the solution at all. OpenAI has said it will not claim the prize.
- Should this change how soon you expect AGI?
- By a little. We moved our own AGI Clock 30 days nearer, because days of sustained autonomous research mathematics ending in a machine-checkable artifact is a real capability jump, and because it reveals an unreleased model ahead of anything on sale, while holding the delta down because nothing has been independently verified and human mathematicians laid the path. Our Immortality Index did not move at all, because no life was extended by it.
Written 16 September 2026, checked 15 and 16 September 2026. Read directly, in full: OpenAI's announcement “On the Navier-Stokes Millennium Prize Problem”, dated 8 September 2026, supplied to us by the owner as a saved page because the site returns 403 to our fetcher, which is the source for every figure and quotation attributed to OpenAI, including the 28 August training start, the rumour of 1 September, the roughly 10,000 agents on Navier-Stokes and about 100 over some 50 hours on Euler, the 88 hours to resolution on 5 September, the 2.7 million messages and about 130 billion output tokens for Navier-Stokes against 4.9 million and about 300 billion across all problems, the 17 hours of Lean formalization by GPT-6 Astra, the account of contacting Alpöge and Buckmaster on 6 September, the dated footnote recording that the “Concurrent work” section was revised on 10 September after an investigation into user inputs, and the refusal to claim the prize; OpenAI's paper “Finite time blowup for Navier-Stokes”, 166 pages, for Theorem 1.1 and its compactly supported smooth force, for the statement that this establishes alternative (C) with (D) following by compact support, and for its related-work section citing Córdoba and Martínez-Zoroa; the companion paper “Finite time blowup for the Euler equation”, 57 pages, for the unforced Euler result; the Lean repository openai/NavierStokesAndEuler, which we downloaded and scanned on 16 September at the state of the 10 September push, giving 2,659 Lean files, about 641,000 lines, Lean 4.34.0-rc2, no sorry placeholders, no declared axioms and no native_decide outside the ComparatorChallenges directory, whose five sorry placeholders are the statements adapted from Google DeepMind's Formal Conjectures formalization of the Millennium problem, together with the repository's own instructions for checking the work with the Comparator tool; Charles Fefferman's official problem statement for the Clay Mathematics Institute, for the four alternatives, for the fact that f must be identically zero only in (A) and (B) while (C) and (D) admit a smooth force satisfying the decay conditions (5) and (9), and for the quoted line about giving reasonable leeway to solvers; Terence Tao's post of 7 September 2026 on finite-time blowup with smooth forcing, for what Alpöge and Buckmaster proved, for the role of Córdoba and Martínez-Zoroa and for his statement that those results do not themselves resolve Navier-Stokes; and the Clay Mathematics Institute's Navier-Stokes page, which still showed the status Active when we checked, alongside its published rules requiring publication in a refereed journal of worldwide repute and two years of general acceptance before the Scientific Advisory Board considers a solution. Taken from reporting rather than read at source: Fefferman's remark that the heroes are Córdoba and Martínez-Zoroa, Clay's president describing the review as deliberately unhurried, Tao's remarks on the strip-mining of open problems, and the reported offer of prize-equivalent money to Alpöge and Buckmaster; principally Quanta Magazine, Fortune, The Next Web and Implicator. Labelled as allegation and counter-allegation, not as fact: Buckmaster's account of being offered reduced credit or the removal of Levent Alpöge's name because of his Anthropic affiliation, and of the remark about ruining his career, against OpenAI's published account of offering a concurrent release and a joint announcement recognising their priority. We could not adjudicate between them and do not try. What our own check does and does not show: scanning a repository is not building it, we did not run lake build, and a static scan cannot detect an unsound tactic or a definition that drifts from the intended meaning, so the absence of a sorry is a necessary condition and not a sufficient one. Corrections to the framing this piece started from: the widely repeated claim that the Millennium statement forbids an external force is wrong for the negative alternatives, and OpenAI's force is compactly supported, which satisfies the stated decay condition outright, so that objection fails twice over; the unforced result is for Euler, not for Navier-Stokes; the problem is about ninety years old measured from Leray and twenty-six measured from the prize; and “AI solved it” remains premature while no independent check has been published. One note on our own instruments: the AGI Clock delta of minus 30 days and the Immortality Index entry of context only are editorial judgments made by TaskNorth, published on those pages with their reasoning, and one commit to change. Disclosure: Levent Alpöge works at Anthropic, whose models this site is built with and whose Claude our knowledge base recommends, and the news collectors written for the AGI Clock and the Immortality Index are built against OpenAI's API, though they are dormant until the owner provisions their credentials.
Trying to work out which AI tools fit your task? Describe the outcome and get a Blueprint: the tools, the prompt, and the steps, with pricing we verified ourselves.