The rumor started the swarm

OpenAI did not wake up one morning and decide to solve a Millennium Prize problem for the poetry of it. It heard a rumor.

On 1 September, according to the company's own write-up, someone whispered that two of the Clay Institute's seven mountain peaks had already fallen. The whisper turned out to be about Tristan Buckmaster at NYU and Levent Alpöge at Anthropic, who were deep into fluid dynamics with the help of large language models, including OpenAI's. Inspired by that rumor, and by a new internal model that OpenAI says is "significantly more capable than GPT-6 Astra," the lab pointed a swarm at every open Millennium Prize problem still on the board.

Eighty-eight hours later, on the order of 10,000 concurrent agents had an analytical proof that three-dimensional Navier–Stokes flow can blow up. Seventeen more hours of Lean formalization, run through Astra, and the company says the logic checks. Across the whole campaign the agents swapped about 4.9 million messages. On Navier–Stokes alone: about 2.7 million messages and roughly 130 billion output tokens. Sébastien Bubeck has put the compute bill in the several-millions-of-dollars range. Sam Altman called it one of the most amazing moments in OpenAI's history this past week.

That is the triumphant version. The version that spent Monday night and Tuesday morning ripping through X is messier, and more important.

The Navier–Stokes equations are Newton's second law applied to a continuous fluid. Aircraft, weather, blood. The Clay question, posed in 2000, asks whether a smooth three-dimensional incompressible flow can develop a singularity: speeds that grow without bound in finite time, even while viscosity tries to smooth things out. Real fluids cannot move infinitely fast. A singularity would mean the continuum model has broken.

OpenAI's solution is a vortex that spirals inward and stretches like spaghetti. The fluid starts at rest. A smooth external force is applied. Kinetic energy stays finite. Velocity still becomes unbounded. The company says this establishes statements "C" and "D" in Clay's official formulation, and it is releasing both a paper and a Lean formalization. It also says it does not intend to claim the million-dollar prize.

Hold that last sentence. Prize politics are not the story. The story is what kind of science this is.

Some independent write-ups overnight still argue about forced versus unforced Clay wording, and about whether "solved" belongs in a headline or only in a press release. Treat the result as OpenAI's claim until the mathematics community finishes chewing. Lean checks a statement. Humans still have to agree that the statement is the one everyone meant.

About twelve hours before OpenAI's Tuesday morning announcement, Buckmaster posted. He and Alpöge had Lean-verified progress on related blowup questions, including forced three-dimensional Euler, the zero-viscosity cousin of Navier–Stokes. Buckmaster's public statement is not gentle. He alleges OpenAI learned of their progress, pressed for coordinated announcements, and floated dropping Alpöge from authorship because he works at Anthropic. Bubeck has called those claims false and inflammatory. OpenAI's published note congratulates the pair on forced Euler, offers them priority there, says it did not see their unpublished materials, and adds the careful hedge that de-identified product-use data cannot be entirely ruled out as having helped its models.

Both AI-enabled efforts sit on top of human scaffolding that Quanta's reporting makes hard to miss. Diego Córdoba and Luis Martínez-Zoroa spent years building analytic "forcing" techniques that do not lean on the dominant computer-assisted approach. Charles Fefferman, who wrote Clay's official problem description, called them the heroes of the story. Martínez-Zoroa's reaction to watching others finish the cascade was almost painfully decent: it would have been nice to do this ourselves, but he was happy for Buckmaster.

That is the emotional center of the week. Not the spaghetti. The handoff.

Fields Medalist Terence Tao's warning, circulating on X overnight, is the part that will outlive the press cycle. When a lab can spin up ten thousand agents the day it hears a rumor, sharing a half-finished hunch stops being a gift to the field and starts looking like a tip to a competitor with a bigger cluster. Open science was always a bet that credit would follow contribution. Agent swarms compress the gap between "someone is close" and "someone else ships."

OpenAI framed the release as evidence that we are in the next period of AI progress, and as a reason to think harder about pacing. That sentence lands differently next to Jakub Pachocki's Sunday essay arguing that maximum-speed scaling is no longer responsible. The same company that wants the world to see how fast the models are moving is asking the industry to invent brakes. Navier–Stokes is the demo. The brake pedal is still a blog post.

And then, because this week cannot leave a crescendo alone, Altman replied to a thread overnight with something close to a dare: you know what, let's try. The suggestion was a room-temperature superconductor hunt on the same swarm playbook. The replies split the usual way. Half the timeline called it more valuable than AGI. The other half reached for replication nightmares and current density. Nobody asked whether the rumor that starts that swarm will come from a physicist's half-finished note, or from another lab's Slack leak.

Lean formalizations will get read. Clay's process, if anyone ever steers toward the prize, takes years of publication and community acceptance. Buckmaster, Alpöge, Bubeck, and OpenAI's denial language will keep trading posts. Córdoba and Martínez-Zoroa's cascade will get the citations it deserves if the field is honest. Tao's worry does not need a resolution to be useful: it is already a design constraint on how mathematicians talk to each other with agents in the room.

The equations may have blown up. The harder singularity is social. A rumor reached a lab that could afford several million dollars of tokens, and eighty-eight hours later the rumor had a Lean file. That is either the new scientific method, or the end of the old one wearing a progress costume. This morning, both readings are still live.