OpenAI reportedly proves finite-time Navier–Stokes blow-up with a 10,000-agent swarm

What was published

On 8 September, according to a detailed write-up on dev.to, OpenAI published a proof that the three-dimensional incompressible Navier–Stokes equations can develop a singularity in finite time. The result, a 166-page paper with an accompanying Lean formalization, constructs a fluid that starts at rest under a smooth external force, keeps its kinetic energy bounded, and yet has a maximum speed that grows without bound as time approaches one. The post frames this as a negative resolution of the Navier–Stokes Millennium Prize problem as the Clay Mathematics Institute wrote it, ninety years after Leray, and reports that OpenAI will not claim the million-dollar prize.

A swarm, not a flagship model

The proof did not come from GPT-6 Astra, the model OpenAI released five days earlier. Per the dev.to account, it came from an internal model, training since 28 August, that OpenAI rates as considerably more capable than Astra, with mathematics among the benchmark areas showing gains. Astra's only role was the final step: producing the Lean formalization and verification in 17 hours.

The machinery above the model was a swarm. The post reports on the order of 10,000 concurrent agents, given a cached snapshot of the internet and code execution, exchanging 2.7 million messages and about 130 billion output tokens across roughly 88 hours from 1 to 5 September. Across all problems the effort attempted, the totals reached 4.9 million messages and 300 billion tokens.

Three orchestration choices stand out. OpenAI reportedly did not bet on a direction, prompting separate groups with statement pairs A and B (solutions stay smooth) and C and D (they can break down). Other groups worked easier problems, and one fell first: the unforced Euler equations, the same question with viscosity removed, resolved by about 100 agents in roughly 50 hours and then fed to the Navier–Stokes groups as a prompt. Agents could talk only within their group, with Codex consolidating the most useful insights between groups into follow-up prompts.

What the theorem actually says

The statement is short. For every positive viscosity there is a smooth external force, compactly supported in space and time, such that the fluid starting at rest under it is smooth before time one, has bounded kinetic energy throughout, and has a maximum speed that goes to infinity as time approaches one. That is statement C of Fefferman's official problem description, and the compact support yields statement D for free. Statements A and B, whether a fluid left alone with no push can break down, remain as open as they were.

The force is the caveat worth being honest about. The post explains that the paper chooses the blowing-up flow first and defines the force as whatever residual the equations leave behind; the whole difficulty is picking a flow whose residual stays smooth through the singular time. A lone background vortex leaves a residual that diverges, and the construction cancels it with spatially oscillatory pulses before removing the remaining errors order by order, and that cleanup is where the paper's length goes. Some mathematicians will call this a resolution of the problem as written, the post argues, others a loophole, and both readings have a point.

Where the fluid stops being a fluid

The blow-up has a shape you can picture: a vortex that spirals inward while stretching along its axis, like water going down a drain. Its core radius shrinks like the square root of the time remaining, while its peak speed grows slightly faster than the inverse of that square root. Long before the singular moment, the core is narrower than the mean free path of air molecules and the speed passes the speed of sound, at which point incompressibility, the assumption underneath the equations, is already false. As the post puts it, the equations are not predicting an infinitely fast fluid so much as announcing where they stop being valid.

The practical consequence for fluid mechanics is, by this account, close to nil. The singularity requires a force engineered to cancel diverging terms at every order at a single point in space and time, and nothing in a wing, a hurricane, or an artery supplies one. Numerical solvers already regularize below the grid scale, and the core's kinetic energy actually decreases toward zero. No engineering code is expected to change.

Why it matters

Three things make this bigger than a prize story. First, the status of the model itself has shifted: the continuum equations were assumed to be a closed, self-consistent description occasionally applied outside its range, whereas now, if the account holds, they are known to contain their own exit, reachable with a smooth push. Second, the method is a landmark in its own right: a research-level theorem produced by a swarm of thousands of agents whose orchestration, from open-ended problem statements to stepping-stone results and machine-mediated consolidation, other teams will study closely. Third, the caveats deserve the same attention as the headline: the unforced case remains open, the prize may go unclaimed, and every figure in this article comes from a single dev.to write-up describing OpenAI's publication rather than from the paper itself. The Lean formalization, once independently checked, would give mathematicians a machine-verifiable way to confirm the result.