OpenAI claims its AI has solved Navier-Stokes

OpenAI claims that an internal artificial intelligence system has solved the Navier-Stokes problem, one of the Millennium Prize Problems in mathematics. According to the company, the proof shows that a fluid can develop a singularity, meaning its speed grows without limit in finite time even though it behaves perfectly smoothly at the start.

The result still has to withstand analysis from the mathematics community. OpenAI has published a proof and a formalization in Lean, a system that translates a mathematical proof into a language a computer can check step by step.

What problem does it solve?

The Navier-Stokes equations describe how fluids move. They are used, among other things, to design aircraft, predict the weather and study blood flow. Their central mystery has remained open for approximately 90 years.

The question focuses on a three-dimensional fluid with constant density and no compression: if it starts moving smoothly, will it remain that way forever, or can a point appear where the equations stop behaving properly?

Viscosity tends to smooth out motion, as when honey slows a swirl. The challenge is to prove whether that force always prevents a mathematical blow-up or whether, under certain conditions, the fluid can concentrate so much speed that the model is no longer valid.

In 2000, the Clay Mathematics Institute included this question among its seven Millennium Prize Problems and offered a one-million-dollar prize for each accepted solution. OpenAI says it does not intend to claim that prize for its result.

A vortex stretched until the model breaks

The solution described by OpenAI starts with a fluid initially at rest and applies a smooth external force. The motion forms a vortex, a swirl that narrows and stretches inward like an increasingly thin strand of spaghetti.

At the center, the region becomes smaller while the speed grows without limit. Even so, the fluid's total energy remains finite. That detail is essential: the singularity does not appear because someone introduces an infinite force, but because of the equations' own dynamics.

The proposal also claims that the terms representing acceleration, pressure, momentum transfer and viscosity grow very intensely, but balance each other precisely. The result would be a smooth external force alongside a speed that shoots up in finite time.

How the AI worked

OpenAI did not use a single chatbot to search for the answer. It organized a system of agents, meaning multiple AI processes that could explore different approaches, communicate and run code.

The group focused on Navier-Stokes brought together approximately 10,000 concurrent agents. These agents:

• Reached the proposed solution in about 88 hours.
• Sent around 2.7 million messages.
• Used approximately 130 billion tokens during the work on Navier-Stokes.
• Needed another 17 hours to formalize and verify the proof with GPT-6 Astra.

The system also produced an answer to a related problem: the regularity of the inviscid Euler equations. In that case, nearly 100 agents worked for about 50 hours to prove a version without an external force.

What changes for you

This does not mean that aircraft, weather forecasts or medical simulations will work differently tomorrow. The equations are already used with numerical approximations and will continue to require them.

The important point is different: an AI has generated a proof in an area where mathematicians have gone decades without a definitive answer, and it has also attempted to express that proof in a format a computer can check. If the community confirms that the argument is correct and fits the official formulation of the problem, it will change how advanced mathematics is researched.

There are also clear limits. A formalization in Lean helps detect logical errors within the system, but experts still need to check that the formalized proof represents the original problem exactly and contains no hidden assumptions. OpenAI's publication is a claim of resolution, not automatic acceptance of the prize.

OpenAI says its internal system is far more capable than GPT-6 Astra and was trained through reinforcement learning on a previously trained model. The company presents the result as evidence of the pace of progress in its models, not as the end of mathematical research.

The open question is not only whether this proof withstands review. We will also need to see whether these systems can reliably explain their discoveries, find new results in other fields and work under controls that allow humans to understand and verify every step. That is where the true reach of this announcement will be measured.