OpenAI reports that a network of AI agents has produced a formal proof suggesting a finite-time singularity in the three-dimensional Navier Stokes equations, a result that could resolve a longstanding Millennium Prize Problem if verified
A network of AI agents assembled by OpenAI has generated a formal mathematical proof that claims to identify a finite-time singularity in the three-dimensional Navier-Stokes equations-a result that, if confirmed, would resolve one of the most persistent open questions in mathematics. The company released the proof and its formalization in the Lean proof assistant, but has not submitted the result for the $1 million Millennium Prize, citing the need for independent expert review.
The core claim centers on a scenario where a smooth, stationary fluid develops a vortex that collapses inward, causing velocity to become unbounded in finite time while the total energy remains finite. This behavior, if it emerges naturally from the equations without external forcing, would demonstrate that the Navier-Stokes equations can break down under their own dynamics, despite the smoothing effect of viscosity. The technical challenge lies in constructing a solution where acceleration, pressure gradients, and viscosity all intensify and balance precisely enough to produce a singularity without introducing infinite external forces.
OpenAI's approach did not rely on a single model or a lucky output. Instead, the company orchestrated a large-scale agent-based search, dividing the problem among thousands of concurrent AI agents operating in isolated environments. Each group pursued different strategies, with some focusing on possible breakdowns and others on regularity. The effort began on September 1, following rumors of progress on major mathematical problems, and peaked at around 10,000 agents. The agents were able to run code and access cached internet resources, with earlier work on the Euler equations guiding the search toward fluid singularities. Codex was used to consolidate promising ideas between agent groups.
After approximately 88 hours, the agent network produced a candidate proof, which was then formalized and checked in Lean by GPT-6 Astra over an additional 17 hours. The process generated 2.7 million messages and roughly 130 billion output tokens. OpenAI has made both the proof and its formalization publicly available, but emphasizes that the result remains provisional until it is scrutinized by the mathematical community. The company also disclosed related work by Levent Alpöge and Tristan Buckmaster on forced Euler equations, but the main focus remains on the unforced Navier-Stokes case.
OpenAI's announcement does not claim the Millennium Prize or assert that the problem is definitively solved. The company frames the result as evidence of advancing mathematical capabilities in machine-directed search, rather than as a final answer. The proof's validity will depend on whether independent mathematicians confirm that the argument meets the rigorous requirements of the official problem statement. If upheld, the result would mark the first time a machine-generated approach has uncovered a mechanism for singularity formation in Navier-Stokes equations-a question that has resisted human analysis for nearly a century.
The scale of the computational effort is notable. Nearly 100 agents spent 50 hours on the Euler equations' regularity problem before the main Navier-Stokes search began. The full Navier-Stokes project involved thousands of agents, millions of messages, and a formal verification step using Lean, reflecting a level of automation and division of labor that would be impractical for human mathematicians alone. However, the process was not fully autonomous: researchers directed the search, selected promising directions, and integrated results between agent groups.
OpenAI's decision to release the proof without claiming the prize or asserting finality is a tactical move that avoids overstepping the limits of current machine reasoning. The company's willingness to expose its methods and outputs to public scrutiny is a positive step for transparency, but the burden of proof now shifts to the mathematical community. Until independent experts confirm the argument, the result remains a demonstration of what large-scale AI coordination can attempt, not a guarantee of what it can achieve. The episode highlights both the potential and the current boundaries of AI-driven mathematical discovery, where automation can accelerate search and formalization but cannot yet replace the judgment and rigor of human verification.
To understand the significance of this claim, it is important to recognize the role of formal proof assistants like Lean in mathematics. These systems allow researchers-and, increasingly, AI agents-to encode mathematical arguments in a way that can be mechanically checked for logical consistency. While formalization reduces the risk of human error, it does not guarantee that the underlying ideas are correct or that all problem requirements are met. The distinction between generating a plausible proof and establishing a result to the satisfaction of the mathematical community remains central to evaluating advances in AI-assisted mathematics.
