OpenAI claims to have solved the Navier-Stokes Millennium Problem with a swarm of thousands of AI agents

On 8 September 2026, OpenAI published a research paper claiming to have solved one of the seven Millennium Problems set out by the Clay Mathematics Institute in 2000: the existence and smoothness of the Navier-Stokes equations. The company says it produced both an analytic proof and a formalization verified in the Lean language, and argues that the result shows that the dynamics of these equations —which describe the motion of incompressible three-dimensional fluids— can develop a finite-time singularity; that is, that fluid velocities can grow without bound despite the presence of viscosity, which normally tends to smooth out motion.

The Navier-Stokes equations, formulated in the 19th century by Claude-Louis Navier and George Gabriel Stokes, model a fluid as a continuous medium rather than tracking individual molecules, and are used in aircraft design, weather forecasting and the study of blood flow. The question that had stayed open for almost 90 years was whether that continuum approximation could break down: whether a motion that begins smoothly could, at some point, stop being so. In 1934, Jean Leray had proved the existence of solutions in a generalized sense, but whether those solutions always remain smooth was still unanswered. According to OpenAI, its system constructed a concrete solution: a vortex that contracts inward and stretches like spaghetti, accelerating in such a way that its energy remains finite at all times, thereby satisfying statements 'C' and 'D' of the problem's official formulation.

The most striking aspect of the announcement is not only the mathematical result, but the method it claims to have used to reach it. OpenAI says it used an internal model —described as 'significantly more capable' than GPT-6 Astra and whose training was still under way— within a multi-agent coordination system with access to tools such as a cached version of the internet and code execution. The agents were organized into groups, able to communicate within each group, and the group that arrived at the Navier-Stokes resolution ran on the order of 10,000 concurrent agents. The company says it maintained the same monitoring and isolation safeguards it applies to all its frontier model evaluations.

The timeline is laid out in detail: training of the new internal model began on 28 August; on 1 September, after hearing rumours that two Millennium Problems had been solved, OpenAI launched an effort to evaluate its system against every open Millennium Problem and other high-impact problems, assigning different variants of the statement (versions that would lead to a proof and versions that would lead to a refutation) to different groups of agents. In parallel, some 100 agents working for around 50 hours solved —to the surprise of the researchers themselves, according to the text— a simpler related problem: the unforced version of regularity for the Euler equations (the zero-viscosity limit of Navier-Stokes), arriving at a refutation. That breakthrough led OpenAI to concentrate resources on Navier-Stokes, redirecting agents from other Millennium Problems and using Codex to consolidate the most useful findings from different groups. The group that ultimately found the solution reached its result on Saturday 5 September, some 88 hours after the first agents were launched; the formalization and verification in Lean took a further 17 hours using GPT-6 Astra. Across the whole effort, the agents sent 4.9 million messages and consumed some 300 billion output tokens; for Navier-Stokes alone, the figure was 2.7 million messages and approximately 130 billion tokens.

The announcement also includes a notable section on concurrent work. The rumour that set off OpenAI's entire effort was, according to the text itself, related to Levent Alpöge, an Anthropic employee, and Tristan Buckmaster, a professor of mathematics at NYU. After completing its project and the Lean verification on 6 September, OpenAI says it contacted both to offer a joint announcement and acknowledge their priority, only to discover that, using an internal Anthropic model, they had in fact solved the forced Euler problem (a different variant from the one solved by OpenAI's system, which was the unforced version). OpenAI states that it had no access to that team's work by any means before its public release, and adds —in an update dated 10 September— that a subsequent investigation confirmed that the Codex prompts used by Buckmaster in the two months before the announcement could not have influenced OpenAI's system in any way, including its training. This passage suggests there were outside doubts or questions about possible data contamination between the two efforts, serious enough to warrant a public clarification.

In closing, OpenAI is explicit in tempering the reach of its own claims: it says it does not intend to claim the Millennium Prize associated with this problem, and frames the result not as a culmination but as a 'snapshot' of progress in AI development. The company links the episode to what it calls 'the next period of progress in AI' and states that one of its central goals is to build systems that are 'steerable, accountable and connected to people', which, it says, may require more deliberate decisions about the pace at which ever-greater capabilities are deployed.

This announcement is best read with the caution usually warranted by extraordinary claims made by the interested party itself: it is a publication by OpenAI about its own non-public system, with no independent peer review of the mathematical result on record in the text beyond the Lean formalization the company says it produced and released for verification. The case is, in any event, relevant for several reasons that go beyond the proof itself: it illustrates an emerging pattern of research through massive swarms of coordinated AI agents (on the order of thousands working in parallel for days), the use of 'cross-pollination' between groups of agents as an exploration technique, and the near-simultaneous appearance of related results produced independently by two rival labs (OpenAI and Anthropic) using their respective internal models, which is in itself a significant data point about the current state of the mathematical capabilities of non-public frontier models.