Artificial intelligence just crossed a frontier that mathematics couldn’t break through for nearly a century.
An internal OpenAI system produced proof that the Navier-Stokes equations can develop a singularity in finite time, solving one of the oldest and most respected problems in modern mathematics. 🤯
If you’re not familiar with that name, don’t worry — we’ll break it all down for you.
The Millennium Prize, offered by the Clay Mathematics Institute, is a list of seven math problems considered the hardest in the world, each carrying a one-million-dollar reward for whoever manages to solve them.
The Navier-Stokes problem has been on that list since the year 2000 and went unsolved for decades, challenging the greatest mathematicians on the planet. The central question of existence and smoothness of solutions for three-dimensional fluids remained open for roughly 90 years.
What nobody expected was that the solution would come from where it did: a multi-agent AI system running approximately 10,000 simultaneous agents, exchanging more than 2.7 million messages over the course of roughly 88 hours of nonstop work.
This isn’t just a scientific milestone — it’s a clear signal that the pace of AI progress is accelerating at a speed few people imagined. OpenAI shared both the full written proof and a formalization done in Lean, a language specifically used to verify mathematical proofs with absolute rigor.
But before diving into the technical details of this historic achievement, it’s worth understanding what exactly was solved and why it matters so much to science, technology, and our everyday lives.
Receive the best innovation content in your email.
All the news, tips, trends, and resources you're looking for, delivered to your inbox.
By subscribing to the newsletter, you agree to receive communications from Método Viral. We are committed to always protecting and respecting your privacy.
What the Navier-Stokes equations are and why they matter so much
The Navier-Stokes equations are a set of mathematical formulas that describe how fluids behave — whether it’s water, air, blood in the human body, smoke rising from a chimney, or even the plasma of a collapsing star. They apply Newton’s second law, the famous F equals m times a, to describe the motion of fluids, treating the fluid as a continuous medium rather than tracking each individual molecule.
They were formulated in the 19th century by physicists Claude-Louis Navier and George Gabriel Stokes, and since then they’ve become the foundation of virtually everything involving fluid dynamics. Engineers use these equations to design more efficient aircraft. Climatologists rely on them to build weather prediction models. Doctors use them to understand how blood flows through arteries. The problem is that, despite all this practical usefulness, nobody could mathematically prove with full rigor whether these equations always produce stable, smooth solutions — or whether at some point they can blow up, technically speaking, generating infinite values that break any computational model.
That breaking point has a technical name: singularity. A singularity, in this context, happens when a fluid’s velocity grows without limit within a finite time interval — meaning the mathematical system simply stops working in a predictable way. Imagine trying to simulate the behavior of a water vortex and, at a certain point, the equation hands you back an infinite number with no plausible physical explanation. That would be a singularity. And what makes it even more striking is that this breakdown would have to happen despite the presence of viscosity, which is precisely the property of a fluid that tends to smooth out motion and slow everything down.
For decades, the big question was: can this actually happen with fluid equations in real life, or is it just a theoretical problem with no practical consequences? Since no real fluid can move infinitely fast, a singularity would mark a failure in the way the equations model the fluid. Beyond that point, you’d need to go back to tracking the behavior of each particle individually to continue describing the system.
A little history helps put the size of this challenge in perspective. In 1934, French mathematician Jean Leray proved that solutions exist in a generalized sense for these equations. But the question of whether those solutions always remain smooth went unanswered and became one of the great obsessions of modern mathematics. In 2000, the Clay Mathematics Institute officially designated the Navier-Stokes existence and smoothness problem as one of the seven Millennium Prize Problems.
The difficulty lies in the fact that the Navier-Stokes equations are highly nonlinear, meaning small changes in initial conditions can produce completely different and unpredictable results over time. This makes it impossible to use traditional mathematical approaches to analyze every possible scenario all at once. It’s like trying to map every drop of an ocean in constant motion. And it was precisely this level of complexity that kept the problem untouchable for so long — even with all the advances in modern mathematics, computing power, and simulation tools available over the past several decades.
How OpenAI’s AI accomplished what mathematics couldn’t do on its own
The approach OpenAI used was radically different from any previous attempt. Instead of putting a single language model to work on the problem, the company developed a multi-agent system where approximately 10,000 AI agents worked collaboratively and simultaneously, each one responsible for a specific piece of the puzzle.
These agents were powered by an internal model described by OpenAI as significantly more capable than GPT-6 Astra. They had access to powerful tools, including the ability to read a cached version of the internet and execute code. The agents were divided into groups that could communicate with each other, and the group that produced the Navier-Stokes resolution involved something on the order of 10,000 agents working at the same time. 🚀
Over the course of roughly 88 continuous hours of processing, the agents involved in this specific problem exchanged approximately 2.7 million messages and consumed around 130 billion output tokens. Each agent contributed a piece of the mathematical puzzle, and the system as a whole kept organizing, verifying, and refining the generated proofs — in a process that somewhat resembles how a large research lab operates, with hundreds of specialists collaborating in parallel, but at an infinitely faster pace.
A curious detail is how it all started. OpenAI says that on September 1st, they heard rumors that two Millennium Prize problems had been solved. Inspired by those rumors and the performance leap of their internal model, the team launched an effort to evaluate the system against all the open prize problems. Different groups of agents received distinct variants of the problem statement, covering both pathways that would lead to a proof and those that would lead to a refutation.
A boost from the Euler equations
Before tackling Navier-Stokes, the system solved a problem considered simpler but still significant: the regularity question for the Euler equations, which are essentially the Navier-Stokes equations without the viscosity term. About 100 agents worked together for approximately 50 hours to produce this refutation, in the version without external forcing applied to the fluid. It was precisely this win that led the team to go all in on Navier-Stokes, redirecting resources and feeding the agents the Euler solution to accelerate their reasoning.
The result was the production of a formal analytical proof indicating that the Navier-Stokes equations can indeed develop a singularity in finite time. The solution they found was a vortex — a fluid whirlpool that spirals inward and keeps getting more elongated, like a strand of spaghetti. This central region shrinks while accelerating, in a way that keeps the system’s energy finite, exactly as the laws of physics require.
The technical challenge was enormous: the terms in the equations describing the motion — such as acceleration, pressure gradients, momentum transfer, and viscosity — all needed to simultaneously become enormous and cancel each other out with extreme precision. This delicate balance is what allows a smooth external force to exist even while the fluid’s velocity grows without bound. In practical terms, the AI found a mathematically consistent scenario where the system of equations collapses because of the fluid’s own motion, not because of some infinite force artificially thrown into the mix.
The formalization and verification in Lean took an additional 17 hours, using GPT-6 Astra. Across all the problems tested, the agents sent 4.9 million messages and consumed roughly 300 billion output tokens. It’s worth noting that the proof will still need to go through the formal peer review process within the scientific community, which is absolutely expected and necessary for any discovery of this magnitude. OpenAI itself made it clear that it does not intend to claim the one-million-dollar prize for this result.
It’s also worth highlighting that this kind of multi-agent system represents a significant evolution in how artificial intelligence is used to solve complex problems. Previously, AI models were mainly used to assist human researchers — offering suggestions, identifying patterns, or speeding up specific calculations. What happened here was different: the system conducted mathematical reasoning autonomously, structured chained logical arguments, and arrived at a conclusion that no human had reached before.
Parallel work that happened at the same time
An interesting chapter of this story involves researchers from another company. The rumor that kicked off OpenAI’s effort was connected to Levent Alpoge, an employee at Anthropic, and Tristan Buckmaster, a math professor at NYU. After completing its project and the Lean verification, OpenAI reached out to the pair to offer a joint release and acknowledge the priority of their work. It turned out that, using an internal Anthropic model, they had produced a resolution of the Euler problem with external forcing applied.
OpenAI recognized the priority of that work and congratulated the researchers on the achievement. It’s worth noting that the proofs are quite different: in the Euler case, Alpoge and Buckmaster proved a result with external forcing, while OpenAI’s system proved a result without external forcing. The company also confirmed, after an investigation, that no specific user data was accessed to solve the problem, and that nothing from the rival team’s work influenced the system.
What actually changes in practice with this discovery
The confirmation that fluid equations can develop singularities has some very concrete practical implications, even if at first glance it seems like a purely theoretical matter. In computational fluid dynamics simulations, engineers and scientists need to know whether the mathematical models they’re using are reliable under all conditions — or whether there are scenarios where they’ll simply fail catastrophically.
If a singularity can form in finite time, that means any simulation software based on these equations needs to include mechanisms for detecting and handling those collapses — something that was previously considered just a theoretical precaution but now carries much more weight. This affects everything from aircraft design to modeling extreme weather events. 🌪️
In medicine, the impact is also significant. Computational models of blood flow are used to plan surgeries, predict cardiovascular risks, and develop devices like artificial heart valves and stents. If the equations governing those models have formal points of failure, medical simulation systems need to be revisited based on this new understanding. On top of that, the discovery opens the door for researchers to develop new versions of the equations, or mathematical extensions that behave more stably under extreme conditions, which could result in significantly more accurate and reliable simulation tools in the coming years.
In the field of artificial intelligence itself, this achievement demonstrates that multi-agent systems have enormous potential for tackling problems that were beyond the reach of any individual approach — human or computational. This should accelerate interest and investment in collaborative AI architectures, where multiple specialized models work together instead of a single generalist model trying to handle everything on its own.
OpenAI itself sees this result as a snapshot of a specific moment, not a finish line. The company states that we’re entering a new era of AI progress and reinforces that one of its core goals is to build steerable, accountable systems connected to people. This may require more deliberate choices about the pace of advancement, as the mission of ensuring artificial general intelligence benefits all of humanity continues to move forward.
The Millennium Prize still has six other open problems, and the scientific community is already speculating whether the same kind of approach could be applied to them. Either way, what OpenAI demonstrated here is that the boundary between what’s humanly possible and what AI can achieve is shifting at a pace that nobody can afford to ignore anymore. 🧠
