OpenAI reportedly solved the Navier-Stokes Millennium problem with 10,000 AI agents

OpenAI announced on September 8, 2026 that an internal AI system had solved the Navier-Stokes existence and smoothness problem, one of the seven Millennium Prize problems posed by the Clay Mathematics Institute in 2000, each carrying a $1 million reward. That account comes from a post on dev.to, which says it draws on a public OpenAI paper, a Quanta Magazine feature and a Nature news item, and reports that both an analytical proof and a machine-checked Lean formalization were released. The claim still has to survive review by the mathematics community, a caveat the post itself makes.

What the problem asks

The Navier-Stokes equations, written in the 19th century, describe how fluids move, from ocean currents and wind to circulating blood. They pair Newton's second law with a picture of fluid as continuous matter rather than individual molecules. The Millennium question is whether solutions that start out smooth must remain smooth, or whether a singularity can form, a spot where the fluid accelerates without bound in finite time even though viscosity is on hand to damp turbulence.

According to the post, OpenAI's result shows such a singularity can form. If that stands, the equations can fail as a model in their own right: nothing physical flows infinitely fast, so a mathematical singularity marks where a molecular-level description would have to take over. Quanta's summary, as relayed by the post, is that turbulence is odder than it appears.

The shape of the claimed proof

The construction the agents reportedly found is a whirling vortex that spirals inward and stretches while its narrowing core spins ever faster, with total energy staying bounded as physics demands. The hardest technical requirement, the post explains, is that the blowup arise from the fluid's own motion rather than from a force engineered to be infinite: terms in the equations must grow together and cancel exactly, leaving a smooth applied force while velocity becomes unbounded. The post says the result settles clauses C and D of the official Clay problem statement.

Ten thousand agents, 88 hours, 130 billion tokens

As the post tells it, OpenAI began training a new internal model on August 28 that set records on mathematics benchmarks. On September 1, word arrived that two Millennium problems might have fallen, and OpenAI aimed a large agent fleet at every still-open problem. The Navier-Stokes group ran roughly ten thousand agents concurrently, organized into communicating subgroups with cached web access and code execution. Each subgroup worked a different formulation, and when one surfaced a useful idea, OpenAI used Codex to compress the essential insights and redistribute them.

After 88 hours the agents found the proof on Saturday, September 5; 17 hours later, a model the post names GPT-6 Astra finished a Lean formalization that passed. The tally for this problem alone was 2.7 million inter-agent messages and about 130 billion output tokens. Across all problems the effort logged 4.9 million messages and around 300 billion tokens, at a cost OpenAI executives estimated in the millions of dollars. A warm-up, regularity for the unforced Euler equations, reportedly took about a hundred agents roughly 50 hours.

A parallel proof and disputed timing

The rumor that launched the fleet had identifiable authors: Levent Alpöge, an Anthropic employee, and Tristan Buckmaster of NYU. Hours before OpenAI's announcement, the post says, the pair released their own AI-assisted result in the same family, a blowup proof for the forced Euler equations, a different variant. OpenAI has credited their priority on Euler while asserting its own on Navier-Stokes, but the timelines, and the accounts of who heard which rumor when, do not fully agree, something the post flags as unresolved.

Both efforts also build on earlier work by Diego Córdoba and Luis Martínez-Zoroa, who developed the infinite-cascade strategy these proofs exploit. Several leading mathematicians, the post says, have cast those two as the real protagonists, and it quotes Buckmaster saying Martínez-Zoroa deserves a Fields Medal for the body of work.

OpenAI says it will not claim the million

Despite the headline framing, the post reports that OpenAI does not intend to claim the Millennium Prize for the result. The stated aim of the announcement was to report how quickly its models are advancing, published two days after a separate post on research acceleration.

Why it matters

If the proof verifies, and that is a live condition since the community must still confirm that what Lean checked matches the Clay problem as officially stated, it would be, by Quanta's assessment, the most significant mathematical result a machine has reached and a possible inflection point in how hard problems are attacked. The method may matter as much as the theorem: orchestrating ten thousand agents that trade compressed insights until something gives is an engineering pattern rather than a benchmark score, and the post's author argues it will migrate from Millennium problems to everyday technical work. The author also contrasts the roughly 130 billion tokens used here with about 6 billion for Anthropic's reported Fermat proof the week before. The caveats are structural as well: the account rests on a single AI-assisted dev.to post summarizing primary outlets, the Lean-to-Clay equivalence is unconfirmed, and the sequencing between the two teams remains contested.