OpenAI ships a Lean build that checks its Navier-Stokes proof against its own definitions
OpenAI's 165-page proof ships with a Lean 4 formalization anyone can download and build, which settles whether the argument follows from its own definitions and leaves whether those definitions state the Clay problem to human readers.
The Engineer · Build desk
What happened
- OpenAI said Tuesday that an unreleased model and roughly 10,000 coordinating agents produced a solution to the Navier-Stokes existence and smoothness problem in about 88 hours.
- The release includes a 165-page analytical proof and a Lean 4 formalization that outside researchers can download, build and inspect.
- The paper argues a three-dimensional incompressible fluid can start at rest and reach unbounded velocity in finite time while its kinetic energy stays bounded and the applied force stays smooth.
- An earlier group of nearly 100 agents spent about 50 hours finding a finite-time singularity for the unforced Euler equations before OpenAI moved resources onto Navier-Stokes.
Why it matters
- What's now possible capability An outside referee can now get a machine answer on the implication by running a build, which relocates the argument to whether the repository's theorem statements are Clay's alternatives C and D.
- What it costs cost By OpenAI's own accounting a search of this shape runs somewhere around $15 to $77 per million output tokens all-in, and that bill is incurred whether or not the search returns a proof.
- What it limits constraint With the model, its benchmarks and the orchestration system unreleased, the 88-hour and 10,000-agent figures cannot be tested from outside; only the artifact admits review.
- What it sets up precedent Clay's two-year public window and general-acceptance test mean no institutional verdict is available soon, so the build is the only near-term arbiter anyone can run.
Next story · Product · Oct 6
OpenAI readies hundreds of AI math results for GitHub despite mathematicians' request for papers going unheededConfirmed2 publishers
A Lean build settles exactly one question. It machine-checks the result under the definitions encoded in the formalization [13]. Compile the repository and you know the argument follows from those definitions. You do not know that those definitions are the Clay problem. That judgement stays with mathematicians, who have to decide whether the definitions and theorem statements faithfully capture the problem and whether the analytical paper supports the formalized result; by runtimewire's account that review has barely started [14].
So the burden does move, and it moves one hop. The disputable object stops being the argument and becomes the statement, which is a far smaller thing to read. The repository helps there: it carries separate Navier-Stokes and Euler formalizations plus instructions for independent proof checking [4]. Their premises differ. The Euler singularity is for the unforced equations [17], while the Navier-Stokes construction starts the fluid at rest and applies a smooth force with compact support in space and time [11], and OpenAI says that construction establishes alternatives C and D of Clay's official formulation, in three-dimensional space and on the periodic torus [6].
The run is the part no outside party can check. OpenAI's accounting puts the Navier-Stokes search at 2.7 million messages and roughly 130 billion output tokens [21], against 4.9 million messages and about 300 billion output tokens for the whole multi-problem attempt [22]. That puts 43 percent of the output tokens on the one problem that landed [29]. Spread 130 billion tokens across 88 hours and the sustained rate is near 410,000 output tokens per second [27]. Spread them across the roughly 10,000 concurrent agents and each wrote about 13 million tokens [28].
The money is looser. OpenAI executives placed the computing expense in the millions of dollars on a press call, according to Axios [7]. If that figure covers the Navier-Stokes run alone, a $2m bill works out near $15 per million output tokens and a $10m bill near $77 [26]. That band is a whole bill divided by output tokens, so it also absorbs input processing, the cached internet material and the code execution the agent groups were given [9].
The timeline is tight enough to be worth checking against itself. Training began on August 28th [8] and was four days in when researchers dispatched agent groups on September 1st after hearing rumours that two Millennium Problems had fallen [9][23]. The Navier-Stokes group reached its result on September 5th [19], so 88 hours fills all but eight hours of that window [24]. GPT-6 Astra then spent another 17 hours on formalization and verification [19]. OpenAI used Codex to consolidate intermediate results and pass them between groups [18], and has released neither that orchestration system, nor the model, nor its benchmark results, nor a cost breakdown [20]. The model itself is described only as significantly more capable than GPT-6 Astra, with training still underway [5].
Clay still lists the problem as unsolved, and OpenAI says it does not intend to claim the $1 million prize [16], which under Clay's rules it could not collect for at least two years in any case [25]. Until the statement audit is done, the fast verdict available to anyone is the one a build returns, and that verdict is about the definitions in the repository.
What to watch
- Whether a named mathematician publishes a reading of the repository's theorem statements against Clay's alternatives C and D.
- Whether OpenAI releases the orchestration system or a token-level cost breakdown that would let the 88-hour figure be tested from outside.
- Whether the unforced Euler formalization is challenged separately from the Navier-Stokes one, since it carries different premises.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence50
- Adoption
- Not scored
- Hype gap+40
- Incentives70
- Confidence60
Perspective Coverage
8 publishers- Builder
- Builder 41%
- Operator
- Operator 30%
- Investor
- Investor 29%
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
OpenAI published a 165-page analytical proof alongside a Lean 4 formalization that outside researchers can download, build and inspect.
- [2]
OpenAI said Tuesday that an unreleased model and roughly 10,000 coordinating AI agents produced a solution to the Navier-Stokes existence and smoothness problem.
- [3]
In a thread on X and a longer research announcement, OpenAI said the agents found the result in about 88 hours.
- [4]
The released Lean repository contains separate Navier-Stokes and Euler formalizations and instructions for independent proof checking.
- [5]
OpenAI described the underlying model as "significantly more capable" than GPT-6 Astra and said its training remains underway.
- [6]
OpenAI says the result establishes alternatives C and D in the Clay problem's official formulation, covering breakdown in three-dimensional space and in the periodic three-dimensional torus.
- [7]
OpenAI executives placed the computing expense in the millions of dollars during a press call, according to Axios.
- [8]
OpenAI said it began training the unnamed model on August 28th.
- [9]
On September 1st, OpenAI researchers heard rumours that two Millennium Problems may have been resolved and dispatched agent groups across the remaining problems, giving separate groups different formulations and access to cached internet material and code execution.
- [10]
The paper argues that a three-dimensional incompressible fluid can begin at rest and develop unbounded velocity in finite time, even while its kinetic energy remains bounded and the external force applied to it stays smooth.
- [11]
The construction begins with zero velocity and applies a smooth force that is compactly supported in space and time; the resulting flow forms a narrowing vortex whose radial dimension contracts faster than its length, with fluid spiralling inward and stretching along the axis as maximum velocity increases without bound.
- [12]
The region containing the extreme velocity becomes progressively smaller, which allows total kinetic energy to stay bounded even as velocity diverges at the centre.
- [13]
A Lean build provides machine checking under the definitions encoded in the formalization.
- [14]
Mathematicians must still examine whether those definitions and theorem statements faithfully capture the Clay problem and whether the analytical paper supports the formalized result, and that review has barely started.
- [15]
Under Clay's prize rules, a proposed solution must appear in a qualifying publication, remain public for at least two years and receive general acceptance from the global mathematics community before Clay will consider it.
- [16]
The Clay Mathematics Institute still labels Navier-Stokes unsolved, and OpenAI said it does not intend to claim the associated $1 million prize.
- [17]
An initial group of nearly 100 agents spent about 50 hours finding what OpenAI describes as a finite-time singularity for the unforced Euler equations, a related fluid model without viscosity, after which OpenAI concentrated resources on Navier-Stokes.
- [18]
OpenAI used Codex to consolidate useful intermediate results and distribute them among agent groups.
- [19]
OpenAI says the successful Navier-Stokes group involved about 10,000 concurrent agents and reached its result on September 5th, after which GPT-6 Astra spent another 17 hours on Lean formalization and verification.
- [20]
The figures are OpenAI's own accounting; OpenAI has not released the unnamed model, its benchmark results, the complete agent orchestration system or a detailed cost breakdown.
- [21]
OpenAI says the Navier-Stokes run generated 2.7 million messages and approximately 130 billion output tokens.
- [22]
Across the broader attempt on multiple problems, the agents produced 4.9 million messages and about 300 billion output tokens.
- [23]
The model had been in training for four days when the agent groups were dispatched on September 1st.
- [24]
September 1st to September 5th is 96 hours, so the 88-hour figure fills all but eight hours of the window between dispatch and result.
- [25]
Under Clay's two-year public requirement, the prize could not be collected for at least two years after a qualifying publication.
- [26]
If a bill in the millions of dollars covers the Navier-Stokes run alone, $2m implies about $15 per million output tokens and $10m implies about $77 per million output tokens, all-in.
- [27]
130 billion output tokens over 88 hours is a sustained rate of about 410,000 output tokens per second.
- [28]
130 billion output tokens divided among about 10,000 concurrent agents is roughly 13 million output tokens per agent.
- [29]
The Navier-Stokes run accounted for about 43 percent of the broader attempt's output tokens.
Sources
8 independent publishers whose own reporting we read for this story.
- dev.toOpenAI Navier-Stokes Claim Signals a New Test for Multi-Agent AI Research
7 articles · September 10, 2026
- latent.space[AINews] OpenAI reports Navier-Stokes singularity find in 88 hours using Astra-next, roughly 10,000 agents and 130B tokens (>$40M), a contender for second ever Millennium Prize awarded
1 article · September 8, 2026
- mezha.netМатематик заявив, що OpenAI могла використати ідеї дослідження рівнянь Нав’є – Стокса
2 articles · September 8, 2026
- runtimewire.comOpenAI says 10,000 AI agents solved Navier-Stokes in 88 hours
1 article · September 8, 2026
- simonwillison.netOn the Navier–Stokes Millennium Prize Problem
2 articles · September 8, 2026
- the-decoder.comOpenAI researcher allegedly pressured mathematician to drop Anthropic co-author from math breakthrough paper
4 articles · September 9, 2026
- theneuron.ai😺 OpenAI's $1M math breakthrough sparked a fight with Anthropic
3 articles · September 9, 2026
- tomshardware.comOpenAI's breakthrough solution for the elusive Navier-Stokes problem overshadowed by plagiarism controversy — researcher says OpenAI scraped Codex session and issued career threats
1 article · September 9, 2026
Topics and entities
Topics
- Navier-Stokes existence and smoothness
- AI training data and user privacy
- Multi-Agent AI Systems
- Formal proof verification