OpenAI's 88-hour Navier-Stokes run leaves provenance in the vendor's own logs
The theorem was checked in Lean, more than most capability claims get, but the part that moves a valuation, that very little human input went in, rests on prompt timestamps and a training-data question that went unanswered.
The Investor · Invest desk
What happened
- OpenAI said an unreleased internal model, run as roughly 10,000 agents from August 28, proved finite-time blow-up for the forced Navier-Stokes equations inside 88 hours, with the argument checked in Lean.
- Days earlier, NYU's Tristan Buckmaster and Anthropic's Levent Alpoge published Lean-formalised blow-up results for 3D incompressible Euler, Boussinesq and porous media, built on Cordoba and Martinez-Zoroa's program.
- Buckmaster says the "very little human input" account gave way on a September 6 call to a whole team, easier warm-up problems including Euler, a prompt itself written by prompting Codex, and an insane amount of compute.
Why it matters
- Who's exposed exposure Anyone underwriting OpenAI is pricing a provenance claim whose only audit trail sits inside the company being underwritten, so the counterparty and the log-keeper are the same party.
- What it limits constraint Because Buckmaster's own reasoning ran through OpenAI's Codex too, outsiders cannot prove or rule out contamination, which pushes priority disputes into records no referee can compel.
- What it costs cost About a hundred agent-years of frontier compute went to a demonstration rather than to serving customers, and the reputational return on that spend is now being argued in public.
Next story · Product · Oct 6
OpenAI readies hundreds of AI math results for GitHub despite mathematicians' request for papers going unheededConfirmed2 publishers
What moves a valuation here is the provenance sentence rather than the theorem. OpenAI describes the producing system as a group of agents running a next-generation model significantly more capable than GPT-6 Astra [3], and Bubeck, on Buckmaster's account, told Alpoge that very little human input went into the result [17]. A Lean file certifies that each step of an argument follows from the last, which is why both camps produced one [2][4]; it records nothing about who chose the line of attack or which prompt started the run.
Price the run on Semafor's account: roughly 10,000 agents from August 28, finished inside 88 hours [16], which is 880,000 agent-hours, call it a hundred agent-years [22], against two mathematicians who spent most of a year making slow progress before things moved from mid-August [15]. The $1m the Clay Mathematics Institute attaches to each of the seven problems [13] does not begin to cover what Buckmaster says he was told was an insane amount of compute [18], so the compute was bought for the demonstration rather than the prize, and it was compute not sold to a paying customer.
Then the dates, which a timestamp would settle in a minute. Eighty-eight hours from August 28 finishes around August 31 or September 1, two or three days before Buckmaster emailed a mathematician at OpenAI on September 3 and got a warm reply that offered him OpenAI compute [23][9]. On the September 6 call, Buckmaster says he asked when the first prompt had been sent, the answer came after a delay, and he eventually placed it after word of his and Alpoge's work had reached OpenAI [6][10]. The public rumour that Anthropic was close was only swirling the weekend before Tuesday's announcement [12], by which point, on Semafor's timeline, the run was already done [16][23]. Both accounts hold only if private word travelled before August 28, and nobody has said when it did.
The contamination question has the same shape, or rather the more interesting version of it: Buckmaster and Alpoge kept their draft proofs inside OpenAI Codex sessions, asked whether the model might have been trained on that data, got no answer, and Buckmaster is careful to say he is not claiming it happened [8]. He was also, by his own account, doing the mathematics with OpenAI's Codex on GPT-5.6 Sol alongside Anthropic's Claude [14]. The party whose capability is in question holds the only records that could clear it.
The counter-thesis earns its place. Mathematics is the one frontier domain where the artifact is checkable by outsiders, Tao called the human pair's result a remarkable achievement [5], Buckmaster himself credits Cordoba and Martinez-Zoroa with the line of attack and says Martinez-Zoroa deserves a Fields Medal [7], and a roughly 100-page Lean-checked proof [11] can be read; if it holds it is only the second Millennium problem ever solved and the first credited to an AI company [24]. Altman says the team wanted a joint release and acted with integrity throughout [21], and Bubeck calls the allegations false and inflammatory while denying that the mathematicians' prompts or proofs were used [20]. The theorem will survive its referees on its own merits. The provenance claim, which is what an underwriter is actually buying, rests on OpenAI's word about its own logs, and if the fuller response Bubeck has promised [20] ships timestamped session records and a training-data attestation covering the days before August 28, that is answered and this reading of it was wrong.
What to watch
- Whether the Clay Mathematics Institute, which pays $1m for the first correct solution to each problem, treats an unreleased model's Lean-checked proof as a claim on the money.
- Whether Alpoge stays on the author list of the Buckmaster paper after the alleged pressure to remove him.
- Whether independent mathematicians, Tao among them, read the roughly 100-page proof and endorse it as complete.
Information, not investment advice. This story is general information about companies and markets, drawn from cited reporting. It is not a recommendation to buy, sell or hold any security; check the cited sources before acting.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence55
- Adoption
- Not scored
- Hype gap+35
- Incentives75
- Confidence50
Perspective Coverage
9 publishers- Builder
- Builder 34%
- Operator
- Operator 27%
- Investor
- Investor 39%
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
OpenAI said on Tuesday that one of its internal models had demonstrated finite-time blow-up for the forced Navier-Stokes equations, one of seven Millennium Prize Problems, each carrying a $1 million reward.
- [2]
OpenAI says the argument was checked in Lean, a proof-verification language used to formally confirm mathematical proofs.
- [3]
OpenAI's post said the proof was produced by a group of agents using an OpenAI next-generation model significantly more capable than GPT-6 Astra.
- [4]
Days earlier, Buckmaster and Levent Alpoge, a mathematician working for Anthropic, published finite-time blow-up under smooth forcing for 3D incompressible Euler, Boussinesq and incompressible porous media, made formal in Lean.
- [5]
Fields Medalist Terence Tao called the Buckmaster-Alpoge result a "Remarkable achievement", and that part of the story is not contested.
- [6]
Buckmaster says that when he asked when the first prompt was sent, the answer came after a delay, and he eventually put it after the news of his and Alpoge's work had reached OpenAI.
- [7]
Buckmaster wrote that the line of attack originates with Diego Cordoba and Luis Martinez-Zoroa, who spent years building forced blow-ups using rough forcing, and said he thought Martinez-Zoroa should be awarded a Fields Medal.
- [8]
Buckmaster says he and Alpoge kept their draft proofs inside OpenAI Codex sessions, that when he asked whether OpenAI's model might have been trained on that data he got no answer, and that he is not claiming it happened, only that the question went unanswered.
- [9]
Buckmaster emailed a mathematician at OpenAI on Thursday, September 3rd, to avoid confusion, emphasising the work with Alpoge was a personal collaboration not supported by Anthropic or NYU; the answer was warm and even offered OpenAI compute.
- [10]
The calls came on Sunday, September 6; Buckmaster says Sebastien Bubeck came on, the two sides had a couple of conversations that afternoon, and Alpoge was not on the line.
- [11]
Buckmaster says he was told an internal OpenAI model had yielded a roughly 100-page proof of forced Navier-Stokes blow-up using smooth forcing along "options c and d" in Charles Fefferman's formal statement of the problem, the narrow route he and Alpoge had secretly selected and that nearly no one else was following.
- [12]
Over the weekend before OpenAI's Tuesday announcement, rumours swirled that Anthropic was on the cusp of announcing that one of its AI models had cracked a Millennium Prize Problem.
- [13]
The Millennium Prize Problems are seven mathematical challenges selected in 2000 by the Clay Mathematics Institute, which offers a $1 million prize for the first correct solution to each.
- [14]
Buckmaster, of New York University's Courant Institute, says he and Alpoge used Anthropic's Claude and OpenAI's Codex powered by the GPT-5.6 Sol model, and used OpenAI's Astra model to help audit and write up results but not for the mathematical reasoning and calculations.
- [15]
Buckmaster says he and Alpoge worked for most of a year making only slow progress, but made rapid progress from mid-August onwards with help from several AI models.
- [16]
OpenAI deployed around 10,000 agents of an as-yet unreleased powerful model on Aug. 28 and solved the problem within 88 hours, proving the equations sometimes break down under extreme conditions.
- [17]
Buckmaster says Bubeck told Alpoge that "very little human input" went into the OpenAI result.
- [18]
Buckmaster wrote that it emerged an entire team had been working on the problem, that this was one of a number of things tried, that work had started on the unforced problem, that the team first set the model on easier problems including Euler, that even the prompt shown to him had been written by prompting Codex, and that an insane amount of compute had been used.
- [19]
Buckmaster alleges Bubeck twice pushed to remove Alpoge as an author due to his Anthropic ties, and once asked him, "Why would you ruin your career?"
- [20]
OpenAI denied the allegations; Bubeck called the claims "false and inflammatory" in a post on X, said he had entered the debate following academic norms, promised a more complete response, and denied using the mathematicians' prompts or proofs.
- [21]
Sam Altman posted that he spent much of the weekend talking with the team who did the work, that Bubeck and everyone else acted with integrity and generosity throughout, and that OpenAI initially believed the other team had also solved the problem and wanted to collaborate on a joint release.
- [22]
Roughly 10,000 agents running for 88 hours is about 880,000 agent-hours, or close to a hundred agent-years of continuous work.
- [23]
An 88-hour run beginning on August 28 completes around August 31 or September 1, two to three days before Buckmaster's September 3 email to OpenAI.
- [24]
If the OpenAI result holds up, it would be only the second Millennium problem ever solved, and the first credited to an AI company.
Sources
9 independent publishers whose own reporting we read for this story.
- cnbc.comOpenAI claims to have solved the 90-year-old Navier-Stokes math problem in 88 hours
1 article · September 9, 2026
- cryptopolitan.comMathematician alleges OpenAI's proof story fell apart on a call
1 article · September 8, 2026
- cryptoslate.comOpenAI just pushed smart-contract security closer to an AI takeover
1 article · September 10, 2026
- decrypt.coOpenAI Says It Solved a $1M Math Problem. A Rival Mathematician Says He Did It First
1 article · September 8, 2026
- fortune.comOpenAI says it cracked one of math’s grand challenges. But there are troubling questions about how they did it—and what it means for us all
1 article · September 8, 2026
- indianexpress.comWhat is the 90-year-old Navier-Stokes problem that OpenAI says AI has solved? | Technology News - The Indian Express
1 article · September 9, 2026
- pivotnews.aiOpenAI says 10,000 AI agents disproved the 90-year-old Navier-Stokes problem
3 articles · September 11, 2026
- qz.comOpenAI AI solves Navier-Stokes Millennium Prize Problem
1 article · September 8, 2026
- semafor.comOpenAI agents find proof to $1 million Millennium Prize Problem
1 article · September 8, 2026
Topics and entities
Topics
- Formal Verification
- Navier-Stokes existence and smoothness
- AI for Mathematics
- Millennium Prize Problems
- AI Training Data Provenance