Picture a whiteboard the size of a city block. At one end, a single equation that has resisted every human attempt to tame it for over a century. And instead of one mathematician standing there with a marker, imagine ten thousand of them, each scribbling in a different corner, passing notes, abandoning dead ends, and handing promising fragments to the next person over. Nobody sleeps. Nobody gets discouraged. That, roughly, is the picture OpenAI painted on September 8, 2026, when it announced that an internal model had produced a proof resolving the Navier-Stokes existence and smoothness problem using something like 10,000 concurrent AI agents.
Navier-Stokes is one of the seven Millennium Prize Problems, the short list of mathematical questions widely treated as the hardest open targets in the field. So the announcement landed hard. And then the arguing started.
What “10,000 agents” actually means
If you read that number and pictured ten thousand separate robots in a server room, close enough. An AI agent, in the sense used here, is a model given a goal and the freedom to take steps toward it on its own: try something, check the result, try again. One agent working alone is a researcher. Ten thousand working at once is closer to a search party.
That distinction matters for how you read the claim. This was not a flash of insight from a single artificial genius. It was brute persistence at a scale no human department could match, aimed at a problem where the hard part is finding a path through an enormous space of possible arguments. Running thousands of attempts in parallel is a strategy, not a miracle, and it tells you something about where AI is currently strong: exploring vast option spaces tirelessly.
Why mathematicians pushed back
The response from the mathematical community was not applause. Twenty-five Fields Medal winners, the field’s most decorated researchers, released a declaration arguing that the race by AI companies to solve famous problems as a benchmark actively harms the science of mathematics.
That is a striking objection, and it is worth understanding on its own terms. Their concern is not that A It is about what happens when the most visible measure of progress becomes a handful of trophy problems. Mathematics advances through a slow accumulation of ideas, definitions, and techniques that get reused for decades. A proof is valuable partly for what it reveals along the way. Treating celebrated problems as scoreboard entries flips the priorities: the win becomes the point, and the understanding becomes optional.
There was also a process question. On September 21, 2026, an independent Advisory Group on Mathematics and AI was formed to advise on the release of these results. Setting up outside oversight two weeks after a major announcement says something about how unsettled the verification question still is. A claimed proof of a Millennium Prize Problem is not true because a press release says so. It is true when other mathematicians can read it, check it, and agree.
The less dramatic story is the more important one
Underneath the headline fight, something quieter has been building. Quanta Magazine has reported on AI models successfully handling research-level questions across various areas of math. A February challenge called First Proof gave entrants one week to have their AI models solve ten research-level problems. By early 2026, according to that reporting, the mood among mathematicians had shifted from shock to something closer to wonder.
Money is moving in the same direction. Renaissance Philanthropy announced 22 grant awards in 2026 for projects in AI for mathematics and theoretical computer science. The mix is telling: future-looking moonshots alongside field-building work, benchmarks and datasets to track progress, and infrastructure. Benchmarks and datasets are not glamorous. They are also exactly what you fund when you want to know whether claims hold up.
What to take away from all this
If you follow AI from the outside, this episode is a useful case study in how to read big announcements:
- Scale is a real strategy. Thousands of agents running in parallel can reach places one model cannot, and that approach will show up in other fields.
- A claim is not a confirmation. Verification in mathematics is a social process, and the advisory group exists because that process has not finished.
- Experts objecting are often objecting to incentives, not capability. The Fields medalists’ declaration is about what gets measured and rewarded.
- The steady progress matters more than the trophies. Research-level problems being solved across many areas of math is the deeper shift.
Ten thousand agents on a whiteboard makes for a great image. But the work that decides whether this moment holds up is happening in the unglamorous places: the benchmarks, the datasets, and the mathematicians reading line by line to see if the proof survives.
🕒 Published: