OpenAI Navier-Stokes result arrives after an 88-hour run by

OpenAI has produced what it calls a resolution of the Navier-Stokes existence and smoothness problem, one of the Millennium Prize challenges posed by the Clay Mathematics Institute in 2000, after an 88-hour effort by roughly 10,000 coordinating agents. The OpenAI Navier-Stokes result, announced on September 8, 2026, comes with a proof, formally checked over an additional 17 hours in the Lean language using GPT-6 Astra, that three-dimensional incompressible fluid flow can develop a singularity in finite time even from a smooth state. OpenAI says it does not intend to claim the $1 million prize attached to the problem.

The run paired an unreleased internal model, which OpenAI describes as significantly more capable than GPT-6 Astra, with tool access that included code execution, a cached copy of the web, and cross-pollination of ideas through Codex. The Navier-Stokes effort alone consumed roughly 130 billion output tokens and 2.7 million messages. Reporting around the announcement has put the computing bill in the millions of dollars.

A companion effort produced what OpenAI calls a disproof of a related Euler regularity conjecture in about 50 hours. Across both problems the program consumed roughly 300 billion output tokens and 4.9 million messages. The agents reached the Navier-Stokes resolution on September 5, three days before the company went public with the work.

The disclosure also carries product-strategy weight. The agent run relied on a system that has not shipped, which means the most capable model OpenAI used is not the one customers can buy. Enterprise buyers watching the announcement are seeing a capability gap between what the company runs internally and what it sells.

What the proof shows

The equations at the center of the claim, named for the 19th-century physicists Claude-Louis Navier and George Gabriel Stokes, model how fluids such as air and water move, and the smoothness question asks whether their solutions stay well-behaved forever or can blow up in finite time. That question has gone unanswered for roughly 90 years. The Clay Institute made it one of seven Millennium Problems in 2000, each carrying a $1 million prize, and six were still open before this announcement.

OpenAI's agents claim the second outcome. Even from smooth initial conditions, a three-dimensional incompressible flow can develop a singularity, a breakdown corresponding to statements C and D of the Millennium formulation. If the proof survives scrutiny, it would be the first Millennium problem resolved by an AI system and, by a wide margin, the most important mathematical result a machine has produced so far.

The scale of the OpenAI Navier-Stokes result

OpenAI's disclosed figures show the two problems as separate runs within one program, and they explain why the effort is being read as an exercise in resource allocation as much as in mathematics.

Run	Time	Output tokens	Messages
Navier-Stokes resolution	~88 hours	~130 billion	~2.7 million
Euler regularity disproof	~50 hours	n/a	n/a
Combined program	n/a	~300 billion	~4.9 million

The coordination design matters as much as the raw totals. Sub-agents worked different parts and variations of the problem in parallel, and OpenAI says promising lines of attack were shared across the swarm through Codex so that progress in one thread could steer the others. At its peak the system used about 10,000 concurrent agents.

Token volumes at that level move the work beyond anything a person can audit by reading. A single problem producing output tokens counted in the hundreds of billions, exchanged across millions of messages, yields more text than any reviewer could examine in a lifetime, so confidence in the proof has to rest on the formal layer rather than on human inspection of the reasoning. The time split points the same way: OpenAI spent 88 hours generating the argument and 17 hours verifying it, making discovery the expensive stage of the pipeline and verification the cheap one.

That inversion is the economic story. For a lab with OpenAI's compute budget, an open research question increasingly becomes a resourcing decision, a question of how many agents to launch, how many tokens to spend, and when to stop. Individual mathematicians cannot bid tokens at that scale, which is one reason the result is being watched so closely.

Verification and the burden of proof

Lean formalization is the strongest check OpenAI offers. A machine-checked proof of statements C and D removes most of the risk that the internal logic is wrong. What human mathematicians still have to judge is whether the formalized statements faithfully encode the original conjecture and whether the construction the agents found answers the problem as posed, and that review is only beginning.

OpenAI also separated generation from checking in a way that is itself notable. The discovery run used the unreleased internal model, while the formalization pass was executed with GPT-6 Astra. That split, if it becomes a norm, would let labs spend their most capable models on exploration and use well-understood models for verification.

Secrecy complicates independent assessment of the OpenAI Navier-Stokes result. The coordinating model has not been released, so outsiders cannot reproduce the run on the same system. OpenAI has published the result together with the Lean formalization, but the discovery process itself, including the paths the agents explored and abandoned across millions of messages, cannot be replayed by the community.

Because the generating system is proprietary and the output is machine-produced, the burden of proof falls on both sides. OpenAI must show that its formal proof matches the claim it is making, and the mathematical community must develop ways to audit proofs at this scale, since reading the reasoning is no longer an option.

OpenAI's decision not to claim the prize removes the most direct test of the rules. The Clay Institute has never had to decide whether a result produced by a proprietary model on a company's own budget qualifies for a prize designed for researchers. By declining, OpenAI leaves that precedent question open while still asserting the mathematical result.

A dispute over provenance

The contested credit predates the announcement. OpenAI says the project was inspired by rumors of concurrent work by the mathematicians Levent Alpöge and Tristan Buckmaster, while maintaining that its own results differ from theirs. Buckmaster has accused OpenAI of cheating and intimidation over the provenance of the result, and OpenAI has acknowledged that it cannot rule out that its agents drew on private research data the mathematician kept in Codex.

Terence Tao has publicly lamented how the episode unfolded. For a field built on careful attribution, the combination of an unreleased model, a multi-million-dollar compute bill, and a disputed data trail makes this a live test of how credit will be assigned when research is driven by agent swarms rather than by named individuals.

Why this matters

The OpenAI Navier-Stokes result is a preview of how discovery will be organized at the largest labs: research questions framed as compute budgets, fleets of agents, and token bills counted in the billions. It also shows that verification, not generation, is the constraint that decides which machine-produced results the wider community accepts and who receives the credit. How mathematicians audit a 130-billion-token argument, and how the provenance dispute is settled, will set the terms for the next wave of machine-driven research.

Sources

On the Navier–Stokes Millennium Prize Problem | OpenAI

Photo by Brecht Corbeel on Unsplash

✔Human Verified

Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.