OpenAI says its 10,000 AI agents have found a way to create a singularity in Navier-Stokes; a claim that has sparked debate and disagreement within the mathematics community.
Think of the swirling of smoke in the air, the turbulence of water behind river rocks, the invisible flow of air around the wings of a Boeing 747, or even the movement of blood in our capillaries; in all these scenes, fluids are in motion. For nearly two centuries, physicists and engineers have described such movements using a powerful family of equations called Navier-Stokes.
The Navier-Stokes equations essentially state the laws of motion for liquids and gases. When we have the initial conditions of a fluid, such as velocity, density, and pressure, these equations show us how the state of the fluid will change in the future. This same mathematical language is used in calculating airflow around airplane wings, water movement, weather forecasting, and many other engineering and physical problems.
Now, suppose that at a given moment, the fluid velocity everywhere has normal and regular values. Do the Navier-Stokes equations guarantee that this state will be maintained in the future? Mathematicians call a solution that has such regular behavior a smooth solution; meaning that a variable does not suddenly spiral out of control over time and reach infinite values.
To tell the truth, no one has ever seen a fluid in a river, a wind tunnel, or amidst the vortices of a storm whose velocity reaches infinity at any point. But for nearly two centuries, mathematicians had not been able to prove that the Navier-Stokes equations themselves are always safe from such a fate.
Audio Summary
Audio summary, created with artificial intelligence
The Navier-Stokes problem asks: "Do smooth solutions always exist for the Navier-Stokes equations in three-dimensional space for all time, or can conditions be found that eventually lead to a singularity?" The importance of the question was so great that the Clay Mathematics Institute included it in the list of seven Millennium Prize Problems in 2000 and set a one-million-dollar prize for its solution.
Interestingly, we are dealing with equations that engineers and physicists use every day, yet one of the most fundamental mathematical questions about the behavior of these equations themselves has remained unanswered.
The Navier-Stokes puzzle asks whether fluid flow always remains regular or if it might reach a singularity in finite time.
In fact, we have been using a tool for decades whose theoretical foundations have not yet been fully proven, and now, there is a possibility that artificial intelligence has found the map to these shaky foundations.
In September 2026, OpenAI claimed that a set of its AI agents had managed to attack the problem from an unconventional path and construct an example that drives Navier-Stokes solutions toward a singularity.
But before we get excited and consider the work of mathematics finished, we must pause for a moment. What exactly does solving Navier-Stokes mean? Does artificial intelligence offer a new, magical formula for predicting the movement of water? Or are we dealing with a deeply mathematical labyrinth behind this grand claim?
First, let's take a look at OpenAI's approach to solving the problem.
How did OpenAI's agents work on the Navier-Stokes problem?
Many famous and unknown mathematicians have spent many years of their lives solving equations, making new discoveries, or developing theories. But on September 1, 2026, an experiment began whose scale and speed were unprecedented in the history of research and bore no resemblance to the routine work of classical mathematicians.
Not a human, nor even a research team at a prestigious university, but a group consisting of about 10,000 OpenAI artificial intelligence agents worked in parallel and simultaneously on one of the most famous and difficult unsolved problems in mathematics.
The agents were not merely executing code that the company's researchers had envisioned; according to OpenAI, they were divided into different clusters and groups like a giant, autonomous research organization. They worked on different ideas, messaged one another, wrote code, critiqued each other's hypotheses, and immediately transferred promising findings from one group to others.
The statistics and figures mentioned for this operation are staggering enough. In just 88 hours, by exchanging 2.7 million intra-network messages and producing a staggering 130 billion tokens, these artificial researchers managed to compile a 165-page proof.
Over the next 17 hours, all these arguments were checked in the Lean proof assistant and received confirmation. OpenAI says its agents have solved one of the Millennium Prize problems, a puzzle that has dead-ended the brightest minds of humanity for decades.
But what is the million-dollar Navier-Stokes problem really, and why has the familiar behavior of fluids frustrated so many mathematicians over all these years? What exactly have AI agents solved?
The million-dollar problem; what is Navier-Stokes really?
Summary: The Navier-Stokes equations describe the motion of liquids and gases. The million-dollar problem asks whether a flow that begins in a smooth state always remains smooth, or if it might reach a singularity in finite time, meaning the fluid velocity in the mathematical model increases without bound. OpenAI says it has created an example that demonstrates such an occurrence.
When you stir a spoon in a cup of tea and pull it out, the liquid swirls around for a few moments, but the small vortex inside the cup gradually weakens, and the surface of the tea quickly calms down because the internal friction of the liquid, or viscosity, dissipates part of the kinetic energy.
Liquids and gases are constantly moving, swirling, colliding, and changing speed. But the behavior of fluids is not always as simple as a cup of tea. In more complex flows, vortices can stretch each other, become narrower, and spin more intensely.
Can viscosity always curb this intensification of motion, or is it possible for the velocity in a small region to exceed any limit? This seemingly simple question leads us to the Navier-Stokes problem; one of the seven Millennium Prize problems, for which OpenAI now claims to have found a solution.
Viscosity is the fluid's brake, but can it prevent infinite velocity?
The Navier-Stokes equations show how the velocity and pressure of a fluid change at any point over time. Just as Newton's second law tells us how forces acting on an object change its motion, these equations also examine how forces such as pressure, viscosity, and external forces change the motion of a fluid.
However, the motion of the fluid itself also makes the calculations more difficult. For example, when a water flow pulls a vortex along with it, the vortex also changes the path and velocity of the surrounding water. The equations must describe this mutual interaction as well.
What factors influence the solution to the Navier-Stokes problem?
To examine the Millennium Prize problem, mathematicians consider a three-dimensional, incompressible fluid; meaning that every part of the fluid maintains its volume while moving. The initial state of the fluid must also be "smooth"; meaning that its velocity and changes must proceed regularly and continuously everywhere, and the flow must not suffer from a singularity from the very beginning.
When we talk about the smoothness of a fluid, we do not mean keeping the fluid stationary or that the fluid has no vortices; sometimes a turbulent and complex flow can also satisfy these conditions.
In the next step, mathematicians must see how much and for how long this smoothness lasts. Do the equations continue to describe the velocity and pressure of the fluid with smooth functions for any arbitrary time in the future? Or is it possible that the flow reaches a singularity in a finite time; a state in which the velocity in a small region exceeds any limit?
Note that a singularity does not simply mean a very high velocity. Whatever number we consider as the speed limit, we will exceed it near the moment of the singularity.
Viscosity acts like a brake and prevents different parts of a liquid or gas from passing by each other at very different speeds; however, the movement of the flow itself can stretch tiny vortices and swirls within the fluid, making their rotation more intense and faster. Can viscosity always prevent a singularity?
Mathematicians must start with a three-dimensional, incompressible fluid whose initial flow is perfectly "smooth".
Of course, when a fluid moves, it drags other parts of the flow along with it, affecting their speed and direction of movement. This effect is called nonlinear advection, a part of the equation where the flow velocity is multiplied by its own rate of change.
Two approaches to solving the Navier-Stokes problem
Mathematicians have two paths to solve the problem. One is to prove that if no external force is applied to the fluid, the flow remains smooth forever, given any smooth start that satisfies the problem's conditions. But the second approach says to construct a flow that is initially smooth, but whose speed exceeds any limit within a finite time.
Even if you find a flow with such behavior, provided the problem's conditions are met, it is sufficient to solve the problem. In this path, one can also use an external force; however, the force itself must remain smooth. You cannot apply an infinite force to the fluid and then attribute its unlimited speed to the behavior of the equations.
OpenAI used the second method to solve the Navier-Stokes equations; in the company's proposed model, there is initially a stationary fluid that is set in motion by an external force, and then the flow concentrates in a small corner, and its velocity tends toward infinity. The external force itself remains smooth throughout this time.
The Mathematics of Singularity; Can Fluid Velocity Reach Infinity?
Summary: OpenAI wants to increase the velocity to infinity without the flow energy also becoming infinite. This goal is achieved by concentrating very high velocity in a small part of the fluid and continuously shrinking that part; as a result, the velocity can become higher and higher, while the total energy remains finite.
Why should a flow move toward infinity and singularity at all? Imagine vortex tubes. When a vortex is stretched in a fluid, it becomes longer and narrower, and its rotational speed increases. Just like an ice skater spinning on ice who increases their rotation speed significantly by pulling their arms into their chest.
In a turbulent or complex flow, large structures pull and amplify smaller vortices and fluctuations, and energy cascades toward finer scales. However, physical intuition alone is not enough to prove a singularity. This same process can concentrate velocity differences in smaller and smaller regions; but the shrinking of the vortex has a significant problem.
As vortices are compressed, the rotational speed of the fluid increases dramatically.
The power of viscosity becomes significantly greater at smaller scales. For this reason, simply shrinking and speeding up a vortex cannot be a barrier against viscosity. Now we must consider how it is possible for the maximum velocity of a flow to approach infinity, while its total energy remains finite?
Energy also depends on the amount of fluid moving at that speed. If the velocity doubles, the energy of that same amount of fluid quadruples; but if the volume of the involved fluid simultaneously reaches one-fourth, the energy does not change.
Perhaps we can better understand this paradox with a visual example. Consider a graph with a bump and assume this bump is compressed; meaning it becomes taller and narrower at every moment. The height of the graph can keep increasing and its width decreasing, but the area under that bump, which represents energy, remains finite.
If we want to translate our hypothetical graph image into the world of fluids, the square of the fluid velocity represents the height, and the fluid volume represents the width of the bump.
Intense concentration is the subtle mathematical trick for overcoming the historical Navier-Stokes impasse.
You probably know that kinetic energy is approximately proportional to the product of the square of the velocity and the volume. So, if the fluid motion is concentrated in a very, very small volume instead of spreading out, the velocity can reach infinity without the total energy of the system reaching the point of mathematical explosion.
In fact, the volume approaches zero, the energy remains finite, but the velocity crosses mathematical boundaries. This intense concentration is exactly the subtle mathematical trick that forms the core of the argument and the alleged counterexample of OpenAI.
The Spaghetti Vortex; Examining a Singularity
Summary: OpenAI designs a vortex that constantly becomes narrower and more elongated. The narrower the vortex becomes, the higher its rotation speed increases, moving closer to the conditions necessary for a singularity. However, the increase in viscosity at small scales can still be an obstacle to this process.
To enable the fluid to reach infinite speed, OpenAI's artificial intelligence designed a very specific physical structure. Consider a cylindrical space with a vertical axis in the center, and the fluid rotates around this axis.
Now, assume the fluid flow is pulled from the surroundings toward this central axis. But as we mentioned earlier regarding the "incompressibility" law of fluids, the fluid cannot accumulate and compress on the central axis. Therefore, when a flow rushes from the surroundings toward the center, it can only move upward and downward along the vertical axis.
As a result of the above process, a vortex is formed that becomes continuously narrower in the radial direction and more elongated along the axis. A structure similar to a very long strand is created; a phenomenon referred to in the descriptions of this research using the tangible metaphor of "spaghetti".
The spaghetti-shaped structure rotates at high speed to overcome the resistive force of viscosity.
Then, a powerful feedback loop is formed. The fluid is pulled inward, and this compression increases the rotation speed (just like that ice skater).
But do not forget that viscosity, or the retarding factor and friction, has more power at small scales and still tries to stop this process. Therefore, the spaghetti vortex alone is not enough to prove the problem.
Battling the Residual: Tiny oscillations that tricked the equation
Summary: Since the main vortex is not an exact solution for the Navier-Stokes equations and creates a residual, OpenAI adds very tiny oscillations to the flow. The nonlinear nature of the equation causes these oscillations (despite having a zero mean) to have a real effect and neutralize the residual; as a result, without the need for an infinite external force, the vortex continues to progress toward a singularity.
The narrowing vortex seems like an ideal model that tends toward infinite velocity. However, in mathematics, when we plug the proposed velocity field into the full Navier-Stokes equations, the equation is not fully satisfied, and the two sides do not equate. In fact, a residual error appears, which mainly manifests itself at the boundary between the vortex core and the laminar flow surrounding it.
To neutralize the residual error, OpenAI introduces families of very tiny waves and loops around the vortex core into the calculations. As time progresses and the core becomes narrower, these oscillations also become smaller and their frequency increases.
The efficiency of these oscillations stems from the nonlinearity of the Navier-Stokes equations. Suppose an oscillation fluctuates between the values +1 and -1; the average of this oscillation is clearly "zero." However, the Navier-Stokes equations are nonlinear, and in a nonlinear structure, values are raised to powers or multiplied together.
The square of +1 and the square of -1 are both equal to 1, meaning an oscillation that should ostensibly have an average of zero produces a real, non-zero effect when it enters the nonlinear system of the equation.
Small waves, by leveraging the nonlinear nature of the equations, delicately restored balance to the system.
The non-zero effect of the oscillations is used to neutralize the residual resulting from the vortex core. Imagine that on one side of a scale is the residual error caused by the vortex core, and on the other side is the effective flow generated by these tiny oscillations.
The residual grows larger as it approaches the time of singularity, and on the other hand, the tiny oscillations create an effect in the opposite direction through the nonlinear part of the equation. The parameters of the oscillations are adjusted so that these two quantities cancel each other out with great precision.
Neutralizing the residual allows the external force required to maintain the equation to remain smooth and bounded, while the vortex core continues to narrow and its velocity tends toward infinity.
Countdown to Singularity (τ): When Time Approaches Zero
Summary: As the moment of singularity approaches (the reduction of the remaining time, or τ), the vortex core becomes more compressed and narrower, but the velocity within it increases sharply. Ultimately, this process allows for the infinite growth of velocity without breaking the boundaries of the system's limited total energy.
Now that we are familiar with the vortex mechanism, let's look at its mathematical structure over time. First, we need to become familiar with a key variable called tau (τ). Tau represents the time remaining until the formation of the singularity and is defined as τ = T - t, where T represents the time of the singularity's occurrence and t represents the current time. Therefore, the closer we get to the moment of the singularity, the smaller the value of τ becomes, eventually approaching zero.
In the structure proposed by OpenAI, the radius of the vortex core decreases proportionally to τ¹ᐟ². The size of the core along the axis also shrinks, but at a slower rate, proportional to τ¹ᐟ²⁻ʰ, where h is considered a small positive number less than one-hundredth.
As time approaches the moment of the singularity, the radius of the vortex core decreases much faster than its length.
As we approach the moment of the singularity and τ becomes smaller, both the radius and the length of the core decrease. However, the radius shrinks faster, and the core, while shrinking, takes on a more elongated and filament-like shape. Simultaneously with the shrinking of the core, the main flow velocities increase. The rotational velocity around the axis and the velocity of movement along the axis grow approximately proportional to τ⁻¹ᐟ²⁻ʰ.
Since the exponent of τ is negative, the closer τ gets to zero, the larger the velocities become, eventually growing without bound. In contrast, the total kinetic energy of the core, which is related to volume and the square of the velocity, changes proportionally to τ¹ᐟ²⁻³ʰ.
When h is small enough, the power of τ remains positive, so as τ approaches zero, the core energy remains finite; but the velocity reaches infinity. Similar to the idea of compressing a graph into an infinitely small region.
In the video below, you can see a visualization of an example of finite-time blow-up for the Navier-Stokes equations.
Euler's Prelude: Why did the agents start with an ideal fluid?
Summary: The agents first turned to the Euler equations (the inviscid version of Navier-Stokes) and, by creating a chain structure, made each oscillation amplify the next smaller oscillation; a process that repeated with increasing speed. Success in this simpler model provided the key insights and tools needed to implement the main Navier-Stokes project.
Solving Navier-Stokes directly, due to the stubborn presence of viscosity, was like climbing a cliff with no hooks or handholds. That is why the AI team chose an easier path to start; what do you think happens if we remove friction from the system?
If we remove the viscosity term from the Navier-Stokes equations, we arrive at the Euler equations, which describe the behavior of a perfectly ideal fluid without any internal friction. In the absence of viscosity, constructing a mechanism that drives velocity to infinity becomes computationally more straightforward.
By removing viscosity, constructing a mechanism that drives velocity to infinity becomes simpler.
According to the project documentation, a group of nearly 100 OpenAI agents were working on Euler simultaneously while investigating Millennium Prize problems, and after this group's success, OpenAI allocated more resources to Navier-Stokes.
The proposed idea for the Euler equations uses a multi-stage chain. The agents started with a large shear flow, where fluid layers slide over each other. They placed a small oscillation on top of this flow.
The main, large flow stretches the small oscillation and amplifies its velocity gradient. After a while, the amplified oscillation itself plays the same role as the shear flow for a smaller oscillation.
The second oscillation is also amplified and, in turn, affects the third oscillation. This process is repeated many times. Each stage occurs smaller and faster than the previous one and provides the necessary conditions for the next stage. Such a chain is called a "self-similar cascade."
In this process, the time interval between generations continuously decreases, to the point where an infinite number of cascade stages occur within a completely finite time interval. Simultaneously, velocity changes become concentrated at smaller scales. In this temporal compression, the maximum velocity gradient has no choice but to tend toward infinity.
The result that the smaller group of agents obtained for the Euler equations was later transferred to the larger Navier-Stokes project. The ideas and techniques resulting from this stage formed the basis for part of the work of about 10,000 agents who were working on the Navier-Stokes problem.
The Lean proof assistant verifies the logic
After the agents completed their reasoning, it was time for formalization. In the Lean environment, every mathematical argument is converted into a set of completely formal and precise steps, and the system checks whether each result is truly derived from the previous premises according to the rules of logic.
Lean verified the logic of the proof in 17 hours
Formalizing 165 pages of proof in 17 hours and receiving verification from this system is, in itself, a remarkable achievement that minimizes the likelihood of common human errors in the chain of calculations. However, the machine verification of the code's logical structure does not mean the proof is accepted by the mathematical community.
The Lean proof assistant tells you whether the reasoning is logically sound, provided your assumptions are correct, but it cannot determine whether the initial assumptions and definitions exactly match the Clay Institute's problem statement. For example, is the AI's definition of smoothness, domain, and initial conditions the same as what the one-million-dollar prize requires?
Why doesn't the Lean proof help mathematics much?
The entry of proof assistants like the Lean program into this matter solves a major problem, but at the same time, it creates a new challenge. We mentioned that a proof assistant can check thousands of logical and tedious steps with unparalleled machine precision to ensure that the final result follows the basic laws of logic to the letter.
For a proof of such gargantuan proportions, such a tool is absolutely vital, as manually checking this very long chain of formulas is beyond any human's capability.
But mathematics is not merely about lining up several thousand correct propositions one after another. An authentic mathematical proof is, in a way, considered a narrative. A good proof must explain why this theorem works, what invisible thread stitches the different parts of the problem together, and most importantly, which technique used in it can be taken and used tomorrow to solve another problem.
It is precisely this meaningful gap that manifests itself in the Navier-Stokes solution project. The formal output of artificial intelligence might be completely logical and flawless for the verification system, but reading thousands of formal steps does not necessarily give the mathematician any clear picture of the deep logic behind the proof. The human mind needs a more compressed version; a transparent narrative that extracts the main mechanism from the dense forest of details.
Lean can guarantee the correctness of a proof; but it does not necessarily explain why that proof works.
This is where the work of mathematicians who were simultaneously working on similar paths enters a vital and new phase. Their new mission is no longer to check the reasoning line-by-line with a red pen. Their work now resembles mining and extraction.
Mathematicians must delve into the heart of this massive formal machine to determine which ideas are truly innovative, which parts or lemmas have independent value, and how this raw volume of logical data can be transformed into a few pure mathematical insights so that other researchers can also benefit from it.
Perhaps the strangest and most shocking consequence of the entry of artificial intelligence into the world of mathematics is this unprecedented division of labor. We have entered a new era, an era in which machines, with their unimaginable search across thousands of parallel paths, dissect the problem; proof assistants guarantee the correctness of these paths; and ultimately, humans must sweat for months or years just to understand what went on inside the machine's mind and what idea was hidden there.
If the solution to the Navier-Stokes puzzle survives the scrutiny, those 88 hours of processing time will be only half the story. The other half of the story relates to the time that the human mind will need to translate the machine's output into digestible and meaningful mathematics.
It is not yet clear whether the Clay Institute will accept the proof
As soon as OpenAI's claim was published, a wave of doubt and discussion swept through the social networks of mathematicians. While public opinion was waiting for the announcement of a historic victory, reference institutions adopted a completely cautious approach. Currently, the Clay Institute still keeps the Navier-Stokes problem on its list of unsolved problems.
For a proof to receive the Millennium Prize Problem label, it must go through various stages; for example, official publication in reputable scientific journals, the passage of a certain amount of time for review by professors and mathematicians, and finally, widespread acceptance in the elite community.
A group of professors and mathematicians have also raised the issue of the originality of ideas, and some researchers have pointed out whether the neural network has truly created new mathematical insight, or has it merely managed, with its stunning processing speed, to assemble previous human papers, techniques, and achievements into a formal format?
Who found the solution first?
The modern path to proving the Navier-Stokes problem began between 2023 and 2025 with the research of two prominent Spanish mathematicians, Diego Córdoba and Luis Martínez-Zoroa; they developed a method for constructing singularities through a cascade of vortices, although they could not reconcile the external force with the stringent conditions of the Millennium Prize problem.
Shortly after, Tristan Buckmaster from New York University and Levent Alpoge from Anthropic continued this same research path with the help of language models and reached results regarding the Euler equations and a version close to the Navier-Stokes problem. News of their progress reached OpenAI, and according to the company itself, it became the motivation for starting the project that eventually led to the Navier-Stokes proof.
The debate is not just about the correctness of the proof; mathematicians have also questioned the origin of the idea and the contribution of the researchers versus the model.
Almost simultaneously with the announcement of OpenAI's result, Buckmaster published a description of his joint work with Alpoge and raised questions about the origin of the new solution. He stated that choosing the version of Navier-Stokes with a smooth external force was not an obvious or clear path that everyone would naturally think to choose.
While advancing their research, Buckmaster and Alpoge had entered their drafts and ideas into Codex and Claude, and Buckmaster asked whether their research data had in any way influenced the company's new model or solution. He, of course, emphasized that he had no evidence for his claim and did not definitively rule out such an occurrence.
In response, OpenAI stated that its researchers and agents had not seen Buckmaster and Alpoge's work before public release, and had not accessed any specific user data to solve the problem. The company did not rule out the possibility that the two mathematicians' data in Codex might have influenced the improvement of the models, although it considered it unlikely.
The ambiguities surrounding the origin of OpenAI's achievement sparked broader discussions about scientific credibility. For example, when researchers provide their unpublished ideas to AI tools, how should the developing company clarify the independence of the solution, the origin of the ideas, and the contribution of each researcher?
From Proof to Understanding; The Dawn of a New Era and Reverse Mathematics
Navier-Stokes was always more than just a formula on a blackboard; these equations were the thin and shaky boundary that connected the abstract world of mathematics to our tangible physical reality. For decades, engineers and scientists have relied on these equations, while their mathematical behavior remained a deep and sealed mystery.
But now, an AI company says its thousands of AI agents have mapped the same dark path that successive generations of mathematical geniuses had failed to find.
However, the fate of this claim or historical proof will be determined not by algorithms, but by human minds. The credibility and precision of the Lean proof assistant will never replace the careful examination of the problem statement, the verification of assumptions, and the understanding of the mathematical meaning of the proposed mechanism.
Even if the scientific community declares this proof one hundred percent flawless tomorrow morning, it will not be the end of the story; it will just be the starting bell for a new mission. We must discover exactly what the machine saw in its processes that had remained hidden from the sharp eyes of humans for centuries.
Perhaps the most exciting and at the same time most shocking chapter of the Navier-Stokes book begins exactly here. We are standing on the threshold of a historical transformation; for the first time in the history of science, machines may have completed the proof before humans have had the chance to understand the deep logic of the argument.
It seems that from now on, the path of scientific evolution will follow a different trend, and perhaps researchers will take a reverse path—a breathtaking journey that no longer leads from understanding to proof, but rather begins with machine proof so that humans can eventually reach "understanding."
This article was updated on October 7, 2026, with new information.
Comments