OpenAI claims Navier-Stokes solution in 88 hours

10,000 AI agents and 88 hours to solve Navier-Stokes

OpenAI has reportedly solved one of the seven Millennium Prize Problems. Or so Sam Altman’s company claims. According to them, their model reached a solution for the Navier-Stokes equations in just 88 hours, utilizing approximately 10,000 coordinated AI agents. The announcement spread like wildfire over two days, triggering three immediate debates: is the proof valid, is it original, and who pays the bill?

Because Navier-Stokes is no academic whim. It is the third most famous Millennium Problem, trailing only Riemann and P versus NP. Whoever solves it wins the $1 million prize from the Clay Mathematics Institute. The rest—the energy costs, hardware requirements, and intellectual credit—is where things get interesting.

What OpenAI announced and why it sounds like a milestone

The news broke via the company’s official account: a "leap" in many benchmarks, ongoing training, and as the cherry on top, the solution to Navier-Stokes by an internal group with 10,000 coordinated agents. All under the usual "strict safeguards": monitoring and isolation.

On paper, this resolves a problem open for nearly two centuries. The Navier-Stokes equations describe how water, air, or blood move; the Millennium challenge asks if these solutions are always smooth or if they blow up at some point. Such a result would have direct consequences for meteorology, aerodynamics, or biomedicine. That is no small antiestéticat.

The original idea comes from two Spanish mathematicians

Here lies the uncomfortable twist. The idea underpinning the result did not emerge from a machine staring into the void: it stems from the work of Diego Córdoba and Luis Martínez-Zoroa, joined by names such as Buckmaster and Alpöge. Terence Tao, one of the field's most respected voices, had already discussed the Spanish researchers' work.

The lingering suspicion is ugly: that the model was fed the developments that the teams themselves published openly for discussion, only to present itself as the author. Some call it outright parasitism. Others respond logically that without that prior human insight, the machine would have had nothing to build upon.

Lean is not AI: verification changes the picture

It is worth separating two things that hype often mixes. Lean, the software used to check proofs, is not artificial intelligence: it is a deterministic formal verifier. It checks whether the logical steps of a proof hold up. If the result passes that filter, it stops being opinion and becomes something verifiable.

Specifically, the finding would not be a proof that all fluids behave well, but a counterexample: a smooth configuration, with equally smooth force and finite energy, that blows up in finite time generating an infinite vortex. The hot button issue is whether that "magic force"—the imp pushing the system—belongs to the theory itself. That is where consensus breaks down.

The $1 million prize and the uncovered light bill

Then there is the money. The Clay Institute prize is $1 million. The cost of training and running models of this caliber, according to circulating calculations, would cover less than one percent of that figure in electricity alone, not counting salaries or machine depreciation.

To put it in perspective: a university department can set up a mini-cluster of three nodes with 1.5 TB of RAM and 512 GB of VRAM for around €150,000. Two professional 48 GB cards go for €14,000. And open models like Qwen-2.5-Math or DeepSeek-Coder fit into much more modest budgets. The argument, inconvenient for the big players, is that with smart programming, you do not need to depend on OpenAI or Anthropic.

Terence Tao and the shortcut that erases the path

Tao has touched the sore spot least advertised: if AI solves these problems by taking shortcuts, we lose all the attempts, tools, and theories discovered along the way. The learning process is worth as much as the result, or more. It is the difference between reaching the summit by helicopter and knowing the route.

The history of mathematics is full of stubborn individuals who actually took the route. Andrew Wiles, locked away for years with Fermat's Last Theorem. Grigori Perelman, solving the Poincaré conjecture and walking away from both the prize and the entire community. That romanticism is exactly what a trained model cannot replicate.

---

If the demonstration withstands formal review, Navier-Stokes will cease to be a Millennium Problem and the headline will be historic. If not, it will remain yet another episode of measured smoke. Given the track record of grandiose announcements, the most likely prediction is that the energy bill will be real regardless of the outcome.