OpenAI says it has solved Navier-Stokes with a swarm of 10,000 agents, but sidesteps the million dollars that would confirm it

By Zendoric · September 9, 2026.

OpenAI claims to have solved, with an internal artificial intelligence system, one of the seven Millennium Problems: the existence and regularity of the Navier-Stokes equations. The Clay Mathematics Institute posed those seven problems in 2000 and offers one million dollars for the formal solution of each one. For context, only one of the seven had been solved until now, the Poincaré conjecture, whose proof was signed by Grigori Perelman, who also turned down the prize, though for very different reasons from those OpenAI cites. The Navier-Stokes equations, formulated in the 19th century from Newton's second law (F=ma), describe the movement of fluids and are the mathematical basis of aircraft design, meteorology and the study of blood flow; according to the company, the question had been analytically open for 90 years.

The specific problem was to determine whether an incompressible three-dimensional fluid can develop a "singularity": a point where velocity grows without limit in finite time and the equations stop describing reality. According to OpenAI, its model proved that it can happen: a fluid initially at rest generates a vortex (a spiraling whirl) that stretches and accelerates until the equation collapses, while keeping energy finite. The proof was formalized and verified in Lean, a language that allows mathematical proofs to be checked step by step mechanically —that is, it verifies that the logical chain is internally consistent, not that the mathematical community has validated that it answers exactly the statement the Clay Institute set out in 2000.

That distinction is the key point OpenAI's statement leaves in the background: the company will not claim the associated million dollars. It says its only goal is to "bring transparency" to progress toward Artificial General Intelligence (AGI, a system capable of matching a human at any cognitive task). As industry context, the Clay Institute's rules require any solution to be published in a peer-reviewed journal and to survive a period of acceptance by the mathematical community before it can qualify for the prize; a corporate press release, however spectacular, does not replace that process.

The deployment behind the result, again according to OpenAI, was enormous: a swarm of some 10,000 AI agents debating and structuring mathematical approaches over 88 uninterrupted hours, with 2.7 million messages exchanged and 130 billion output tokens consumed on this problem alone. The system used was not one of its usual commercial platforms, but an experimental model in training since late August that, the company claims, far exceeds the capabilities of GPT-6 Astra; the final formalization and verification of the proof required an additional 17 hours of algorithmic work by GPT-6 Astra itself. It is a figure that says as much about the result as about the method: raw compute applied to millions of coordinated attempts, not a flash of human intuition.

The timing of the announcement is no accident either. According to the source, the race to crack the challenge accelerated in early September on rumors that researchers from Anthropic and New York University had reached a similar result. The company qualifies that its rivals solved a different problem, the Euler equations (which describe fluids without viscosity, a simplified version of Navier-Stokes), and maintains that its own proof is independent and free of external data. Two labs claiming a landmark mathematical milestone in almost the same week, each qualifying what the other solved, is a pattern we already recognize: the competition to narrate leadership toward AGI is starting to set the pace of what gets published and when, more than the certainty of what is published.

That context explains why The Verge, which covered the same announcement, headlined its piece by noting that controversy surrounds OpenAI's milestone. We do not know the full detail of that controversy, but the headline alone is telling: the community does not appear to have received the announcement with the unanimity the company's statement conveys.

Our reading, in the short term, is one of deliberate caution. We have spent a long time watching AI labs announce "historic mathematical results" that dilute once they pass through review by the expert community: formalizations that do not quite capture the original statement, or simplifications presented as the full problem. A proof being verified in Lean guarantees internal logical consistency, not that it answers the question Clay posed. And OpenAI's decision not to claim the prize, wrapped in the language of scientific modesty, can also be read as a way of announcing the result without yet exposing it to the scrutiny that collecting it would require.

In the long term, however, that this is even plausible —that two rival labs are competing to attack, with swarms of agents, a problem open for 90 years and with direct applications in aviation, meteorology and medicine— is exactly the kind of progress that supports our underlying thesis: AI as an engine of accelerated scientific discovery, one more step toward a science that solves in months what once took generations. If these tools mature with peer review as a filter and not as an obstacle to dodge, the coming decade could deliver more verifiable advances in mathematics and physics than the last century. The risk is not in the ambition of the announcement, but in the race between labs ending up rewarding the speed of publishing over the certainty of having solved what it claims to have solved.