10,000 AI Agents Solved a 200-Year-Old Math Problem in 88 Hours. The Bill Was $15 Million.
OpenAI says an unreleased model cracked the Navier-Stokes Millennium Prize problem. The numbers behind the announcement say each agent-hour cost about $17, the $1 million prize now buys one-fifteenth of a single attempt, and the two-year verification clock, not compute, is the bottleneck that matters.
Seventeen dollars and five cents. That is what one hour of frontier AI mathematical labor appears to cost, once you divide OpenAI’s numbers the way nobody at the press conference did. On September 8, the company announced that an unreleased internal model had solved the Navier-Stokes existence-and-smoothness problem, one of the seven Clay Millennium Prize Problems, after as many as 10,000 AI agents worked the problem for about 88 hours. At the same briefing, OpenAI said a customer wanting to run the same problem would pay around $15 million. Fifteen million dollars divided by 880,000 agent-hours is $17.05 an hour, roughly what a mid-tier cloud GPU rents for, except this GPU spent its hours doing mathematics that had resisted humans for two centuries.
What the model claims to have found is a blowup: a scenario in which a perfectly smooth fluid at rest develops a singularity, a vortex that shrinks while spinning faster until its speed hits infinity in finite time. Put less mathematically, the equations describing water and air can, on paper, explode. Because no real fluid can reach infinite speed, the result suggests the Navier-Stokes equations are not always a reliable mirror of physical reality. OpenAI published a 165-page proof plus a machine formalization that checks the argument step by step, and mathematician Luis Martínez Zoroa, who did seminal work on the problem, called it “a truly remarkable result” in Nature.
Speed matters here as much as the result. This was a seven-day sprint from rumor to announcement, the fastest competitive response in the problem’s history:
| Date | Event |
|---|---|
| Sep 1 | Rumors reach OpenAI that two mathematicians solved a version of the problem with Anthropic’s models; OpenAI focuses its prototype on Navier-Stokes (Nature) |
| Sep 3 | Tristan Buckmaster (NYU) emails an OpenAI mathematician for clarification (his statement) |
| Sep 6 | Two calls with OpenAI’s Sébastien Bubeck; Buckmaster is told of a ~100-page forced-blowup proof built with “an insane amount of compute” |
| Sep 7 | Buckmaster and Levent Alpöge post three stepping-stone results (porous media, Boussinesq, 3D Euler); Anima Anandkumar (Caltech) posts a zero-viscosity solution |
| Sep 8 | OpenAI announces the full solution: 165 pages, 10,000 agents, ~88 hours, ~$15M |
Now run the economics, because the press releases left the most interesting numbers on the floor. At $15 million for 165 pages, each page of proof cost about $90,909. Against the $1 million Millennium Prize, the attempt cost fifteen times the bounty, and OpenAI says it will not claim the money anyway (phys.org). When the Clay Institute set the prize in 2000, $1 million was roughly twenty years of a senior mathematician’s salary, a life-changing bounty designed to pull talent toward the hardest questions. Today a single attempt costs fifteen prizes. As an incentive, the bounty is finished; capability signaling between labs replaced it.
There is a second calculation hiding inside OpenAI’s phase breakdown, and it is the more important one. Break down the named phases: 1,000 agents for 50 hours on the Euler warmup and 10,000 agents for 11 hours on the extension, 160,000 agent-hours combined (New Scientist). But the headline campaign totals 880,000 agent-hours. Only about 18 percent of the compute went into the two successful phases. Everything else was search: OpenAI had been testing its prototype on all six unsolved Millennium Problems, plus failed attempts, plus the machine formalization. Discovery, it turns out, is mostly search, not the final run. Anyone pricing “one more Millennium problem” at $15 million should understand that most of the bill pays for the misses.
Then there is the dispute, which deserves telling straight because both sides are on the record. Buckmaster and Alpöge spent a year extending a forced-blowup program originated by Diego Córdoba and Luis Martínez-Zoroa, using Anthropic’s Claude and OpenAI’s Codex. They hit their breakthrough August 15 and verified it in Lean on August 22. Buckmaster’s statement says that on the September 6 calls, OpenAI admitted its first prompt had been sent only days earlier, after word of their work reached the company, and declined to answer whether the model had been trained on their Codex sessions, where they had stored a year of drafts as paying customers. He also describes two rejected proposals for coordinating the announcements and an exchange in which he was asked, “Why would you ruin your career?” OpenAI categorically denies using their proof or prompts, says its Euler proof differs from theirs, and chief research officer Mark Chen said no humans accessed Codex work in progress, adding of the thousands of agents, “that’s also our understanding.” Buckmaster is careful about what he does not claim: “I have not seen OpenAI’s proof. I do not know what their model did, or how. I do not know whether our data was used. I am not accusing anyone of anything.”
Zoom out and the week looks less like an upset than a culmination. In May an OpenAI model cracked a decades-old Paul Erdős conjecture. Then Claude Fable 5 found a counterexample to the Jacobian conjecture, open for nearly a century. Last week an AI formalized Fermat’s Last Theorem in 11 days. A year ago, expert forecasters gave AI roughly a 20 percent chance of solving a Millennium problem by 2030 (WSJ). Those forecasters were not slightly wrong; they were four years early on a binary question. That calibration failure should haunt every timeline estimate currently circulating about AI capabilities.
The Strongest Case Against This Article
Here it is at full strength, and it starts with Terence Tao: “There’s been this very strange and unprecedented decoupling, this year alone, between getting answers and getting understanding” (New Scientist). No human has read the 165 pages. A machine formalization verifies that each step follows, but it cannot verify that the definitions encoded at the start faithfully capture the intended mathematics; a wrong definition, formally verified, is still wrong. Independence remains genuinely unresolved, and the discrepancy between the ~100-page forced-blowup proof described to Buckmaster and the 165-page full solution announced publicly has no public explanation. Moreover, this is a negative existence result. It does not make weather models better, wings safer, or blood-flow simulations more accurate. It says the equations can break, which working engineers already suspected. And the incentive structure now rewards landmark-hunting over understanding: Buckmaster, who lived it, calls this a “Deep Blue–Kasparov moment” and argues the community needs “serious and unhurried discussion” about what mathematics is for, rather than a race to feed famous conjectures into a $15 million furnace. If the proof collapses under refereeing, every number in this article’s first paragraph becomes a monument to premature arithmetic.
What This Does Not Prove
Honest bounds on everything above. I have not read the 165-page preprint; every claim about its contents comes through press reporting. That $15 million figure is OpenAI’s own press-conference estimate for a hypothetical customer rerun, not an audited invoice, and the agent counts and hours come from the same briefing. Because the model is unnamed and unavailable, nobody outside OpenAI can reproduce the result today, which means the most basic scientific check, independent replication, is currently impossible. My referee math assumes 5 to 10 pages a day of careful reading and a small qualified referee pool; those are estimates, and the Clay’s two-year acceptance clock only starts after peer-reviewed publication, which has not happened yet. Buckmaster’s account is one side of a disputed story, and OpenAI’s denials are equally unverified. Treat the $17.05 as arithmetic on reported numbers, not as a measured cost.
What You Can Do
Four moves, depending on who you are. If you follow the story: ignore headlines until referee reports exist. Under the Clay’s rules, a solution needs peer-reviewed publication plus two years of community acceptance before the institute even convenes a committee (USA Today), which makes 2028 the earliest possible verdict. Watch the preprint’s revision history, not the press cycle. If you design prizes or grants: denominate bounties in compute, not dollars. A $1 million prize that costs $15 million to attempt is decoration. If you run a math department or fund research: budget for Lean and formalization pipelines now, because machine-checked proof is becoming the standard of evidence and the departments without that infrastructure will be unable to evaluate, let alone produce, this kind of result. And if you build AI systems: study the Euler warmup ladder. OpenAI did not aim 10,000 agents at the hard problem on day one; it solved the easier cousin with 1,000 agents, then scaled. That laddering pattern, warmup then scale-up, is the template to expect in every field where AI starts hunting landmarks.
The Bottom Line
A $15 million, 165-page, machine-checked proof now claims one of mathematics’ seven hardest problems, produced in 88 hours by agents costing about $17 an hour, in a race so tight that rival results landed within a day of each other. Whether the proof survives is now a question for referees, and there are not enough of them: at careful reading speed, three referees need 50 to 100 working days just to get through the pages, before the Clay’s two-year clock even starts. Compute stopped being the binding constraint on mathematical discovery sometime this week. Human verification bandwidth is the bottleneck now, and nothing in the announcements suggests anyone has a plan to scale it.