OpenAI has put a proposed solution to one of mathematics’ most famous problems on the table: a 166-page Navier–Stokes paper, accompanied by public Lean proof files. They reportedly spent $22 million in compute, 6 days, and 10,000 agents to solve the problem, and it looks so far to be money well spent.
That gives mathematicians something concrete to inspect. It also gives everyone watching AI a much more interesting story than another benchmark score.
If the result holds up, it would show AI systems helping produce a proof at the literal frontier of mathematics. The accompanying dispute over credit raises a separate question: How should discovery work when the companies supplying researchers with tools are also pursuing discoveries themselves?
Both deserve attention. First, the math, which is incredible.
What “solving Navier–Stokes” actually means
The Navier–Stokes equations describe fluid motion. The famous existence and smoothness problem asks whether their solutions can keep behaving well indefinitely under specified conditions—or whether a smooth beginning can lead to a mathematical breakdown.
The official problem statement offers several routes to a solution. Two involve proving that smooth solutions always exist without external forcing. Two involve demonstrating breakdown, with a smooth external force permitted.
So “solving” the problem can mean proving that a failure is possible. It does not have to mean finding a formula that predicts every swirl in the ocean.
OpenAI’s paper constructs a three-dimensional flow that starts at rest and develops unbounded velocity in finite time, while its total kinetic energy stays bounded. A smooth force acts on the fluid.
The proposed mechanism concentrates motion into a shrinking vortex. Speed rises as the intense motion occupies less space. The hard part is making the fluid dynamics produce that behavior while keeping the applied force smooth.
Viscosity—the internal friction that tends to smooth differences in motion—is included. Establishing breakdown despite that smoothing effect is central to the claim.
The paper identifies its result with alternatives C and D in Clay’s formulation.
The forcing distinction is essential. Those alternatives explicitly allow it, so its presence is not automatically a technicality that invalidates the result. But this construction does not establish the same outcome for Navier–Stokes with no external force. Those are different statements.
Nor does a mathematical singularity mean real water can move infinitely fast. The result concerns the limits of a mathematical model. It does not make every existing fluid simulation useless or deliver an instant shortcut to perfect weather forecasts.
There is also an Euler result—and it is different
OpenAI separately released a paper on the unforced Euler equations, which describe an idealized fluid without viscosity.
That paper claims finite-time breakdown starting from smooth initial data, without an external force. Here, the theorem concerns the growth of velocity gradients: how sharply velocity changes across space.
Keep those two results straight: forced Navier–Stokes and unforced Euler.
That distinction also matters to the credit dispute, because the work by NYU mathematician Tristan Buckmaster and Anthropic researcher Levent Alpöge concerned forced Euler.
The research effort behind the announcement
OpenAI’s account describes roughly 10,000 concurrent agents in the successful Navier–Stokes group, using an internal model more capable than GPT-6 Astra. The effort began September 1; a resolution came after about 88 hours, followed by 17 hours for Lean formalization and verification.
The Navier–Stokes work consumed about 130 billion output tokens. Researchers redirected agents toward promising results and consolidated insights between groups. The Euler result helped guide the subsequent effort.
That description matters when evaluating what the achievement demonstrates. It combines model capability, substantial computation, tools, coordination, and research decisions.
A headline about AI solving a famous problem compresses all of that into two letters.
The broader implication is that research capacity may increasingly depend on how effectively people organize machine effort: choosing questions, allocating resources, combining partial results, and checking the outcome. The useful comparison for a research organization is the cost and reliability of that entire process.
The Neuron’s coverage of OpenAI’s move toward automated research explores that same shift. More capable agents can expand the work researchers attempt, while making oversight and evaluation more consequential.
A released proof changes the validation question
The public repository contains formalizations for both results, build instructions, and instructions for independent checking with Comparator.
That is a meaningful step toward scrutiny. Readers can examine a specific claim and its supporting artifacts.
Lean checks formal mathematical reasoning against precisely defined statements and assumptions. This makes it particularly valuable when AI produces arguments that would otherwise require readers to trust long, convincing-looking explanations.
But there are several layers of confidence. A successful formal check establishes a result within its formal setup. Experts still need to assess whether that setup matches the intended mathematical problem and understand what the theorem actually says.
Formal correctness, human understanding, and prize recognition are related, but distinct.
Clay’s rules require publication in a qualifying outlet, at least two years after publication, and general acceptance in the mathematics community before it considers a proposed solution. OpenAI says it will not seek the prize.
The credit dispute remains unresolved
According to Scientific American’s reporting, Buckmaster and Alpöge’s work built on an approach developed by Diego Córdoba and Luis Martínez-Zoroa. The pair obtained an AI-assisted Euler result in August.
Buckmaster says they used private Codex sessions during the work and questioned whether their unpublished material contributed to OpenAI’s effort.
He also alleges that OpenAI’s Sébastien Bubeck proposed a publication arrangement excluding Alpöge because of his Anthropic affiliation. Buckmaster says that when he threatened to go public, the response included: “Why would you ruin your career?”
These are Buckmaster’s allegations as reported by Scientific American, not established findings of misconduct.
OpenAI says it contacted the pair after completing its project, offered a joint announcement and access to its prompts, and recognizes their priority on forced Euler.
It denies seeing their work before publication or accessing specific user data for the solution. It nevertheless cannot rule out de-identified usage data helping improve its models.
That acknowledgment does not establish that the pair’s research entered training. The denial also should not be stretched into a claim that every possible data connection has been ruled out.
The public record leaves questions about provenance and the parties’ interactions. A proof checker cannot resolve those questions.
What this signals for AI-assisted research
Our take: the released papers deserve to move the scientific achievement to the center of the story. The dispute deserves careful reporting alongside it.
A valid proof can remain valid even if its credit is contested. That's not a new, or even uncommon, situation. A compelling account of independent discovery still needs evidence. Neither issue should become a substitute for examining the other.
For research teams and businesses, this points to a practical requirement for the next phase of AI adoption. Valuable work needs a record: which materials entered the process, what people contributed, what the system produced, and how the result was checked.
For AI providers, that record is part of earning trust. Customers bringing unfinished ideas into a tool need confidence in the institution behind it as well as the model.
OpenAI has released work that could represent an extraordinary mathematical milestone. The next step is to establish what the proofs settle—and give a clear account of the people, ideas, and systems that made them possible.
Related reading
- OpenAI Chief Scientist on AI Recursive Self-Improvement — How automated research changes the pace of AI development and the demands on oversight.
- GPT-6 Astra: Everything You Need to Know About OpenAI’s New Model — Context on OpenAI’s public model and the expanding role of agents.
