In September 2026 came an extraordinary claim: an internal OpenAI system had solved the Navier–Stokes Millennium Prize Problem.
Almost immediately, Tristan Buckmaster published a statement describing closely related work with Levent Alpöge, raising questions about priority, unpublished material entered into AI systems, and even OpenAI–Anthropic rivalry.
The headlines reduced this to:
AI solved Navier–Stokes, and mathematicians accused it of stealing their work.
But the real story is more interesting.
Before asking who deserves credit, we must ask:
\[\boxed{\text{What exactly was the Navier–Stokes problem that was solved?}}\]
Only then do the mathematics, the history, and the controversy become clear.
1. What the Navier–Stokes Millennium Problem Actually Asks
Start with the actual fluid equation
For an incompressible fluid in three spatial dimensions, the Navier–Stokes equations are
\[\partial_t u +(u\cdot\nabla)u = \nu\Delta u-\nabla p+f,\qquad \nabla\cdot u=0,\]
with initial condition
\[u(x,0)=u_0(x).\]
Here \(u(x,t)\) is the velocity field, \(p(x,t)\) is the pressure, \(\nu>0\) is the viscosity, and \(f(x,t)\) is an externally applied force.
Charles Fefferman's official formulation for the Clay Mathematics Institute says explicitly that \(u_0\) and \(f\) are prescribed data, whereas \(u\) and \(p\) are the unknown fields. Setting \(\nu=0\) gives the Euler equations. Clay Mathematics Institute: Navier–Stokes formulation.
This is important because Navier–Stokes is not merely an abstract mathematical problem. At heart it is a problem in fluid mechanics.
The nonlinear term \((u\cdot\nabla)u\) allows the flow to transport and amplify its own structures, whereas \(\nu\Delta u\) describes viscous diffusion and tends to smooth them.
The great question is whether viscosity is always powerful enough to prevent a smooth three-dimensional flow from concentrating itself into a singularity.
What does “global smoothness” actually mean?
Local smooth existence is not the mystery.
If the initial velocity \(u_0\) is sufficiently regular, standard PDE theory gives a smooth classical solution for at least some interval
\[0\leq t<T.\]
The question is whether the solution can always be continued indefinitely:
\[0\leq t<\infty.\]
That is what “global” means here: global in time.
A global smooth solution requires \(u,p\in C^\infty\) for all finite times, together with an appropriate finite-energy condition such as
\[\int_{\mathbb R^3}|u(x,t)|^2\,d^3x<C.\]
Fefferman notes that if the maximal classical existence time \(T\) is finite, then the Navier–Stokes velocity necessarily becomes unbounded near \(T\). This is not a behavior that Fefferman imposes by definition. It follows from continuation theory.
If everything remained sufficiently bounded as \(t\to T^-\), local existence theory would let us continue the solution past \(T\). But \(T\) was assumed maximal. Therefore something must lose regularity.
\[\text{regularity remains controlled near }T\Longrightarrow\text{solution extends beyond }T,\]
and hence
\[T<\infty\Longrightarrow\text{regularity breakdown}.\]
Fefferman's official formulation.
Fefferman did not formulate one Millennium statement, but four
This point is essential, because much of the public discussion has confused the official Millennium problem with its most famous version.
| Alternative | Space | Force | Required conclusion |
|---|
| A | \(\mathbb R^3\) | \(f=0\) | Every smooth initial flow has a global smooth finite-energy solution. |
| B | Periodic \(T^3\) | \(f=0\) | The analogous global-smoothness result. |
| C | \(\mathbb R^3\) | Smooth \(f\) allowed | Construct smooth data for which no global smooth finite-energy solution exists. |
| D | Periodic \(T^3\) | Smooth \(f\) allowed | Analogous finite-time breakdown. |
Alternatives A and B are existence-and-smoothness theorems. Alternatives C and D are counterexample problems. This is the conceptual key.
C is asking for a counterexample
Let us write the universal proposition explicitly:
\[G:\quad \text{Every admissible smooth }(u_0,f)\text{ produces a global smooth finite-energy solution.}\]
Symbolically,
\[\forall (u_0,f),\qquad \exists\,u_{\mathrm{global,smooth}}.\]
Alternative C asks you to show that this universal proposition is false. It is enough to find one particular pair \((u_0^\ast,f^\ast)\) such that
\[\boxed{\text{there is no global smooth finite-energy solution for these data.}}\]
So C really is:
\[\boxed{\text{construct a counterexample to universal global regularity.}}\]
But C is not the negation of A
Alternative A says \(f=0\) and asks whether every smooth initial velocity remains smooth forever. Alternative C allows \(f\neq0\). Therefore both A and C could logically be true.
We could eventually discover that
\[\boxed{f=0:\ \text{every smooth Navier–Stokes flow remains regular}}\]
while simultaneously
\[\boxed{f\neq0:\ \text{some very carefully designed smooth driving causes blowup}.}\]
There would be no contradiction.
So the question most physicists have traditionally associated with the Navier–Stokes problem—can an unforced smooth three-dimensional viscous fluid spontaneously form a finite-time singularity?—remains open.
OpenAI's claimed result establishes C and D, not A or B. OpenAI, Finite Time Blowup for Navier–Stokes.
Thus we should distinguish
\[\boxed{\text{The official Millennium formulation has apparently been resolved}}\]
from
\[\boxed{\text{the famous unforced regularity problem has been resolved}.}\]
The second statement is false.
The Clay Mathematics Institute has used cautious language, saying that the problem has “apparently been settled” and stressing that verification and attribution will be deliberately unhurried. Clay Mathematics Institute announcement.
2. What OpenAI Claims to Have Constructed
The paper is entitled Finite Time Blowup for Navier–Stokes. Strikingly, the author printed on the paper is simply
\[\boxed{\text{OPENAI}}.\]
Its Theorem 1.1 states that for every \(\nu>0\) there exist a smooth compactly supported force \(f\), and smooth velocity and pressure fields \(u,p\) defined for \(0\leq t<1\), satisfying
\[\partial_tu+(u\cdot\nabla)u-\nu\Delta u+\nabla p=f,\qquad \nabla\cdot u=0,\]
with the remarkably simple initial condition
\[u(x,0)=0.\]
Thus the fluid begins completely at rest.
The solution has bounded kinetic energy,
\[\sup_{0\leq t<1}\|u(t)\|_{L^2}<\infty,\]
but
\[\limsup_{t\to1^-}\|u(t)\|_{L^\infty}=\infty.\]
Consequently there can be no globally smooth finite-energy solution with the same initial data and the same force. That last conclusion is exactly Fefferman C.
Finite energy does not mean finite maximum velocity
How can \(|u|\to\infty\) while kinetic energy remains bounded? Because kinetic energy is an integral:
\[E=\frac12\int |u|^2\,d^3x.\]
A function can become arbitrarily large on an arbitrarily small region while its integral remains finite.
A crude example is a sequence of velocity fields whose magnitude is \(U_n=n\) inside a region of volume \(V_n=n^{-3}\). Then
\[\|u_n\|_\infty=n\to\infty,\]
but
\[\|u_n\|_2^2\sim U_n^2V_n=n^2n^{-3}=n^{-1}\to0.\]
Infinite height can be compensated by shrinking width. This is essentially what happens dynamically in the claimed Navier–Stokes solution.
The shrinking-vortex scaling
Define \(\tau=1-t\). As the blowup time approaches, \(\tau\to0^+\).
The OpenAI construction has a shrinking vortex core with radial size
\[\ell_r\sim\tau^{1/2}\]
and axial size
\[\ell_z\sim\tau^{1/2-h},\qquad 0<h<\frac1{100}.\]
Therefore
\[V_{\rm core}\sim\ell_r^2\ell_z\sim\tau^{3/2-h}.\]
The characteristic large velocity components scale as
\[U\sim\tau^{-1/2-h},\]
so \(U\to\infty\).
Now compute the kinetic energy carried by this core:
\[E_{\rm core}\sim U^2V_{\rm core}\sim\tau^{-1-2h}\tau^{3/2-h}=\tau^{1/2-3h}.\]
Since \(h<1/100\), we have \(1/2-3h>0\). Thus
\[\boxed{E_{\rm core}\to0}\]
while simultaneously
\[\boxed{U_{\max}\to\infty}.\]
The singularity is not an explosion of total energy. It is an extreme concentration of motion.
Why does the singularity occur at \(t=1\)?
There is no spatial boundary involved. The “boundary” is simply the end of the smooth time interval.
The elementary function
\[g(t)=\frac1{1-t}\]
is perfectly smooth for every \(t<1\), but \(g(t)\to\infty\) as \(t\to1^-\). It therefore cannot be smoothly extended through \(t=1\).
The same idea applies to the velocity field. For every fixed \(t<1\), \(u(x,t)\) is smooth. But along points approaching the spatial origin, the velocity becomes arbitrarily large as \(t\to1^-\). The choice \(t=1\) is just a convenient normalization of the finite blowup time.
Why “limsup”?
The theorem states
\[\limsup_{t\to1^-}\|u(t)\|_\infty=\infty.\]
This means that arbitrarily close to \(t=1\), the velocity norm reaches arbitrarily large values. Equivalently, there are times \(t_n\to1\) such that
\[\|u(t_n)\|_\infty\to\infty.\]
The norm need not diverge monotonically.
The apparently strange role of the force \(f\)
Normally we think of \((u_0,f)\) as the input and solve for \((u,p)\). So why does the construction appear to do the opposite?
Because C is a counterexample problem. We are free to construct the bad input.
For a divergence-free velocity \(u\) and pressure \(p\), define the residual
\[R[u,p]=\partial_tu+(u\cdot\nabla)u-\nu\Delta u+\nabla p.\]
Then set
\[f=R[u,p].\]
The Navier–Stokes equation automatically holds. But that alone proves nothing, because for a generic singular \(u\), the resulting \(f\) will also be singular.
The difficult requirement is
\[\boxed{u\text{ becomes singular while }f\text{ remains }C^\infty.}\]
The construction adds carefully designed oscillatory and correction fields so that the singular pieces of the residual cancel and the final force extends smoothly through the blowup time.
The logical construction therefore runs backwards:
\[u,p\longrightarrow f.\]
But once \(f\) has been obtained, freeze it. Call it \(f_*(x,t)\). Then the physical initial-value problem is again
\[(u_0=0,f_*)\longrightarrow u,\]
and the already-constructed \(u\) is an exact solution.
What exactly is the counterexample contradicting?
Not Navier–Stokes itself: the singular velocity field satisfies Navier–Stokes.
Not conservation of energy: the total kinetic energy remains bounded.
And not alternative A: A concerns \(f=0\).
The counterexample contradicts the universal proposition
\[\boxed{\text{smooth initial conditions + smooth forcing}\Rightarrow\text{global smooth classical flow}.}\]
The claimed result says that this implication is false.
There exists a smooth force, acting on a fluid initially at rest, for which the classical velocity field becomes singular in finite time.
Do they actually write \(u(x,t)\) down?
Yes and no.
The theorem is written existentially: there exist \(u,p,f\). But the proof is constructive.
This is not a nonconstructive argument that merely says some mysterious solution must exist. On the other hand, the answer is not a simple elementary formula such as
\[u(x,t)=\frac{x}{1-t}.\]
The solution is built from self-similar profiles, similarity coordinates, vector potentials, oscillatory pulses, correction fields, cutoffs, and an iterative hierarchy designed to cancel the singular residual.
So the right description is
\[\boxed{\text{constructive existence, but not a simple closed-form solution}.}\]
And \(p\) is indeed the pressure.
What has and has not been solved
OpenAI claims a constructive proof, backed by Lean formalization, that smooth forcing can drive initially resting Navier–Stokes flow to finite-time blowup while kinetic energy remains bounded. If correct, this establishes Fefferman C and D and resolves the official Clay formulation.
But it does not settle the unforced problem:
\[f = 0 : \text{global smoothness versus finite-time singularity} .\]
Alternatives A and B remain open. Thus the Millennium formulation may be resolved while the most famous Navier–Stokes regularity question remains unanswered.
3. From Human Insight to Industrial-Scale AI Mathematics
Euler versus Navier–Stokes
Euler is
\[\partial_tu+(u\cdot\nabla)u=-\nabla p+f,\qquad\nabla\cdot u=0,\]
whereas Navier–Stokes is
\[\partial_tu+(u\cdot\nabla)u=\nu\Delta u-\nabla p+f.\]
So
\[\boxed{\text{Euler}=\text{Navier--Stokes with }\nu=0.}\]
The extra viscous term \(\nu\Delta u\) is not a minor correction. It continually smooths small-scale velocity structure. This is why the jump from Euler blowup to fully viscous Navier–Stokes was decisive.
Before OpenAI: Córdoba and Martínez-Zoroa
The OpenAI result did not appear in an intellectual vacuum.
Buckmaster himself says that the basic idea of the entire program belongs to Diego Córdoba and Luis Martínez-Zoroa. Buckmaster's statement.
In 2023 Córdoba and Martínez-Zoroa constructed finite-time blowup for forced three-dimensional incompressible Euler. Their forcing was regular but not \(C^\infty\); roughly speaking, the mechanism arranged successive amplification of increasingly concentrated vortex structures. Córdoba–Martínez-Zoroa, arXiv:2309.08495.
The conceptual mechanism was already present:
\[\text{large-scale strain}\to\text{amplification at smaller scales}\to\text{concentration}\to\text{singularity}.\]
Hypodissipative Navier–Stokes: approaching viscosity gradually
Córdoba, Martínez-Zoroa and Fan Zheng then pushed the method from Euler toward a dissipative fluid equation.
Ordinary Navier–Stokes contains \(-\nu\Delta u\), which damps high-frequency modes like \(|k|^2\). A hypodissipative model uses weaker fractional dissipation, for example \(|\nabla|^\alpha u\) with \(\alpha<2\).
Ordinary Navier–Stokes corresponds, in this notation, to \(\alpha=2\).
Córdoba, Martínez-Zoroa and Zheng obtained forced finite-time blowup for small positive \(\alpha\), below approximately \(0.093\). Córdoba–Martínez-Zoroa–Zheng, arXiv:2407.06776.
Conceptually, the route was
\[\text{Euler}\longrightarrow\text{weak dissipation}\longrightarrow\boxed{\text{ordinary Navier--Stokes}}.\]
Where Boussinesq enters
Another neighboring system is the Boussinesq equation. Schematically,
\[\partial_tu+(u\cdot\nabla)u+\nabla p=\theta e_z,\]
together with an evolution equation for a scalar field \(\theta\), often interpreted as temperature or density variation, and \(\nabla\cdot u=0\).
The scalar produces buoyancy. Physically, this is the sort of coupling involved in convection. Mathematically, Boussinesq provides another laboratory in which vorticity amplification and singularity mechanisms can be tested.
It is not one of Fefferman's alternatives. It is a neighboring physical PDE in which the singularity mechanism was developed and tested.
Buckmaster and Alpöge enter the story
Tristan Buckmaster is a professor of mathematics at NYU's Courant Institute whose research focuses heavily on singularity formation in fluid equations. His earlier work with Vlad Vicol on nonuniqueness of weak Navier–Stokes solutions was recognized with a 2019 Clay Research Award. Tristan Buckmaster, NYU.
Buckmaster and Levent Alpöge took the Córdoba–Martínez-Zoroa program as their starting point and used several LLM systems extensively to push it further.
Buckmaster says their collaboration was personal rather than an institutional Anthropic project. They used Anthropic's Claude as well as OpenAI tools, including Codex. Buckmaster's statement.
Their major advance was essentially
\[\text{rougher forcing}\longrightarrow\boxed{\text{smooth forcing}}.\]
By August 15 they had smooth-forced blowup results for Boussinesq and three-dimensional incompressible Euler. They subsequently verified the Euler argument in Lean on August 22. Buckmaster also reported that they believed they had a hypodissipative Navier–Stokes result, though it was not yet ready for release.
They had therefore not solved Fefferman C, but they had moved very close to the relevant frontier.
What the OpenAI system then did
OpenAI says that on September 1 it heard rumors that major open mathematical problems had been solved and decided to test a new internal model on all the remaining Millennium problems and several related problems. Separate agent groups were assigned versions A, B, C and D. OpenAI's account.
The company also assigned agents easier neighboring questions. One of those was Euler.
Approximately 100 agents worked for around 50 hours and produced what OpenAI describes as an unforced Euler blowup result:
\[\nu=0,\qquad f=0.\]
That is in one respect stronger than Buckmaster–Alpöge's Euler theorem, because their Euler result used external forcing. But Euler is not a Millennium Prize Problem.
At this stage humans made an important strategic decision. OpenAI says:
“Once we saw the Euler solution, we thought that Navier–Stokes was the most promising problem.”
Resources were shifted away from the other Millennium problems, and the Euler result was supplied to the Navier–Stokes agents. Groups were cross-pollinated using Codex to consolidate useful intermediate discoveries.
The Navier–Stokes effort involved on the order of \(10^4\) concurrent agents. OpenAI reports approximately \(2.7\) million messages and \(130\) billion output tokens for the Navier–Stokes part alone.
The claimed solution was reached after approximately 88 hours, followed by another 17 hours of Lean formalization and verification. OpenAI's account.
This is not the ordinary picture of a mathematician using ChatGPT. It is something closer to industrialized mathematical search.
Human versus AI is the wrong picture
Both sides were already hybrid.
Buckmaster and Alpöge represent
\[\boxed{\text{AI-integrated academic mathematics}},\]
while OpenAI represents
\[\boxed{\text{industrial-scale AI mathematics}}.\]
Humans chose the problems and directed the search; AI systems performed much of the detailed exploration and proof production.
So the real contrast is not human versus AI, but two different scales of human–AI mathematical research.
What is Lean?
Lean is not a fluid simulator, and it is not Mathematica.
Mathematica is mainly for symbolic and numerical computation. Lean is a formal theorem prover: it translates definitions, assumptions, and proof steps into precise logic, then checks them with a small trusted kernel.
A useful comparison is
\[\text{Mathematica}\approx\text{powerful calculator},\]
while
\[\text{Lean}\approx\text{logical compiler + extremely strict referee}.\]
Lean verifies that the formal proof follows from the formal statement. Humans must still verify that the formal statement captures the theorem they actually intended.
4. From Mathematical Breakthrough to Scientific Controversy
The strange authorship: “OPENAI”
The 166-page paper lists a single author:
\[\boxed{\text{OPENAI}}.\]
That is unusual. OpenAI wants to emphasize the role of its AI system, but there was still a human research team behind the work.
Unlike “ATLAS Collaboration,” which names an identifiable scientific collaboration, “OpenAI” names an entire corporation.
This raises a new problem for scientific authorship:
\[\text{contribution}+\text{credit}+\text{responsibility}.\]
AI-generated research makes all three harder to assign.
Now the controversy
There was no completed Buckmaster–Alpöge Navier–Stokes proof for OpenAI to copy. But OpenAI acknowledges that rumors of their progress helped trigger its September 1 search.
Buckmaster’s concern was that smooth forcing was a very specific route already being pursued by his group, while drafts had also been entered into Codex. OpenAI says an internal investigation found those prompts could not have influenced the result; Buckmaster says he does not know whether any data were used and makes no accusation.
So the real controversy is about priority, data trust, and attribution — not a proven theft of a proof.
The more troubling part may be corporate competition
According to Buckmaster, OpenAI later proposed that he present the Navier–Stokes result while Alpöge was excluded, with Alpöge’s employment at Anthropic cited as a complication.
If that account is accurate, the issue goes beyond plagiarism: a personal mathematical collaboration became entangled with
\[\boxed{\text{OpenAI versus Anthropic}.}\]
Corporate rivalry had entered scientific attribution.
Who deserves credit?
The story is too complex for a single-author narrative:
\[\boxed{\text{Córdoba + Martínez-Zoroa}\to\text{Córdoba + Martínez-Zoroa + Zheng}\to\text{Buckmaster + Alpöge}\to\text{OpenAI + large-scale AI system}.}\]
Córdoba and Martínez-Zoroa developed the forced-blowup route; Zheng helped extend it to weak dissipation; Buckmaster and Alpöge pushed it to smooth-forced Euler using LLMs; and OpenAI claims the final step to ordinary Navier–Stokes, giving Fefferman C and D.
The long-term historical attribution is still unsettled.
The Deep Blue–Kasparov moment
Buckmaster called this a “Deep Blue–Kasparov moment,” recalling IBM’s 1997 defeat of world chess champion Garry Kasparov.
The analogy is striking because frontier mathematics was supposed to require open-ended conceptual invention, not merely search. Yet this episode suggests that large-scale reasoning models, interacting agents, and formal verification may now be entering that territory.
The difference is crucial: humans opened the conceptual path and chose the targets. What AI changed was the scale and speed with which the nearby mathematical possibilities could be explored.
What I think is really new here
The superficial story is:
\[\text{AI solved a famous problem}.\]
The deeper story is that mathematical research may have acquired a new industrial scale.
Buckmaster and Alpöge represent one model:
\[\text{a few mathematicians}+\text{powerful AI collaborators}.\]
OpenAI represents another:
\[\text{human direction}+10^4\text{ reasoning agents}+\text{massive computation}+\text{formal verification}.\]
And these new modes immediately collided with old questions of priority, authorship, unpublished work, and institutional competition.
The broader lesson
This is neither simply “AI replaced mathematicians” nor “AI merely copied human mathematics.”
Human mathematicians opened the conceptual path; researchers working with LLMs accelerated it; and an industrial AI system appears to have crossed a major remaining gap, with formal verification helping check the result.
What changed is deeper: AI entered the research process itself — not merely as a calculator or search tool, but increasingly as a collaborator. That immediately raises new questions of priority, authorship, responsibility, unpublished ideas, and corporate ownership.
The Navier–Stokes episode may therefore be remembered not only for fluid mechanics, but as the moment the old boundaries between mathematician, tool, collaborator, institution, and author began to blur.
That may be the real Deep Blue–Kasparov moment.
