10,000 AI Agents Cracked a 90-Year-Old Math Problem in 88 Hours — But Who Deserves the Credit?

Every time you check a weather forecast, fly on an airplane, or have blood pressure measured, you are relying on the Navier-Stokes equations — the fundamental mathematical description of how fluids move. Derived nearly two centuries ago, these equations underpin weather modeling, aircraft design, ocean current simulation, and cardiovascular medicine. But for over 90 years, mathematicians have been unable to answer a deceptively simple question about them: can their solutions blow up?

On September 8, 2026, OpenAI claimed the answer is yes — and that its AI found it in 88 hours.

The $1 Million Question

In 2000, the Clay Mathematics Institute designated the Navier-Stokes existence and smoothness problem as one of seven Millennium Prize Problems, each carrying a $1 million reward. The challenge: prove whether smooth solutions to the three-dimensional Navier-Stokes equations always remain well-behaved, or whether they can develop singularities — points where fluid velocity rockets toward infinity.

For 26 years, no one could settle it. Then OpenAI threw 10,000 AI agents at the problem.

How GPT-6 Astra Solved It

The system behind the breakthrough is GPT-6 Astra, OpenAI's most powerful model deployed in a multi-agent configuration unlike anything previously attempted in mathematics. Roughly 10,000 autonomous agents worked concurrently for 88 hours, exchanging between 2.7 and 5 million inter-agent messages and generating approximately 130 billion output tokens. The computational cost: several million dollars, 5.2 megawatt-hours of electricity, and an estimated 28,000 kilograms of CO2 emissions.

The result was a 166-page mathematical paper plus a formal proof written in Lean, a programming language that allows every logical step to be machine-verified. An additional 17 hours of computation produced the Lean formalization, making it one of the most rigorously checked proofs in mathematical history.

The AI didn't start from scratch. It built on "infinite cascade" techniques developed by mathematicians Diego Córdoba at the Instituto de Ciencias Matemáticas in Madrid and Luis Martínez-Zoroa at CUNEF Universidad — a method for combining sequences of non-singular solutions to produce singularities. Charles Fefferman of Princeton, who wrote the Clay Institute's official problem statement, called Córdoba and Martínez-Zoroa the "heroes" of the breakthrough.

What the Proof Actually Shows

The specific finding is both elegant and strange. The proof identifies a vortex configuration — a spinning swirl of fluid spiraling inward like spaghetti being wound on a fork — in which fluid velocity approaches infinity while total energy remains finite. This is an extraordinarily delicate mathematical balance: a finite-time blow-up, proving that smooth Navier-Stokes solutions can develop singularities.

There are important caveats. The singularity occurs at approximately 70 nanometers — a scale far below where the continuous-fluid assumption holds in the physical world. This means the result doesn't change how engineers model airflow over a wing or blood through an artery tomorrow. And the proof applies specifically to the case with smooth external forcing applied to the fluid. The no-external-force case — what many mathematicians consider the "pure" version of the problem — remains unsolved.

The Credit Controversy

But the mathematics quickly became secondary to a fiercer argument about who deserves the credit.

Approximately 12 hours before OpenAI's announcement, NYU mathematician Tristan Buckmaster released a statement — partly on behalf of Anthropic researcher Levent Alpöge, who had been working on the related Euler equations as a personal project — alleging that their unpublished advances had been leaked to OpenAI. Buckmaster claimed the leak may have inspired the prompts used to guide GPT-6 Astra's agents, and further alleged that OpenAI may have drawn on interaction data from his and Alpöge's use of OpenAI's Codex tool.

OpenAI conducted an internal investigation and denied misusing Codex logs. But the company did make a notable admission: the work was "inspired by rumors" of the competitors' progress. OpenAI offered Buckmaster coauthorship on the press release — but excluded Alpöge, a detail that further inflamed the dispute.

Buckmaster was unsparing. He described the AI-generated proofs as "the most horrendous I have ever read" and called one generated paper "AI slop." The controversy echoes a pattern we've seen before — the ongoing battle over AI companies and intellectual property is no longer limited to music and publishing. It's now playing out at the highest levels of mathematics.

OpenAI has stated it does not plan to claim the $1 million Clay Prize.

What's Still Unsettled

Despite the fanfare, the Clay Mathematics Institute has not officially accepted the result. The problem remains listed as unsolved on the Institute's website. Clay has acknowledged the problem has "apparently been settled" but noted that its review process is "deliberately unhurried" — their rules require publication in a qualifying outlet plus at least two years of community acceptance before awarding the prize. If accepted, it would be only the second Millennium Prize Problem to be resolved, after Grigori Perelman's proof of the Poincaré conjecture in 2003.

The Lean formalization allows machine-checking that accelerates verification, but it doesn't replace the community vetting process. Mathematicians need to understand the proof, not just confirm its logical steps.

The Bigger Question: AI as Mathematician

The Navier-Stokes result arrives at a moment when the relationship between AI and mathematics is already fraught. In the weeks following the announcement, 26 Fields Medallists — recipients of math's highest honor — issued a joint statement warning that treating historical problems as AI benchmarks threatens mathematics as a discipline.

Fields Medallist Terence Tao captured the tension precisely: "The actual solving of these problems is only a proxy goal for the primary goal of developing mathematical understanding and insight." In other words, the answer matters less than the understanding — and a 166-page machine-generated proof that humans struggle to read may provide the former without the latter.

This concern isn't theoretical. We've already seen what happens when AI agents operate at scale without adequate oversight. The question isn't whether AI can solve hard problems — September 2026 proved that it can. The question is whether the way agentic AI is reshaping work will preserve the human understanding that made these problems worth solving in the first place.

Why It Matters

The Navier-Stokes result is a landmark, but it's a complicated one. It settles a foundational question about turbulence — confirming that fluid velocities can, mathematically, blow up into infinity — and it demonstrates that AI has moved from computational assistant to independent problem-solver on humanity's hardest unsolved challenges.

But the credit dispute, the environmental cost, and the mathematical community's deep unease about AI-as-benchmark all complicate the narrative. This isn't simply "AI solved a famous math problem." It's the opening chapter of a much harder question: what does mathematical achievement mean when the achiever is a swarm of 10,000 agents that can't explain its own reasoning?

The proof says fluid velocity can spiral to infinity — but only at a scale of 70 nanometers, far smaller than the molecules that make up actual fluids. It's a mathematical truth that may never touch the physical world. And the debate over who found it, and how, is only getting started.