AI News

OpenAI says tens of thousands of agents solved the Navier-Stokes equations in 88 hours, forgoing the $1 million prize

Phoenix TechnologyUpdated Nov 16 07:24

AI summary

OpenAI says that an as-yet-unreleased new model of its drove about ten thousand agents to solve the Millennium Prize Problem, the Navier-Stokes equations, within 88 hours; afterward, GPT-6 Astra used Lean to complete proof verification, and the $1 million prize was forgone. The agents exchanged 2.7 million messages among themselves and output about 130 billion tokens, and Altman estimates the compute cost of this run at about $10 million. The first half of another Millennium Prize Problem, Yang-Mills, was released as Lean code on GitHub by a five-person review panel, and the review document states that parts such as the mass gap have not yet been proven.

Why read this

AI score 90

The original text recounts the process by which OpenAI used tens of thousands of agents to solve the Navier-Stokes equations within 88 hours, and provides details on cost and subsequent review.

Topics

Tags

Body · AI translation

10,000 AI clones crack a Millennium Prize Problem in 88 hours! Altman went two nights without sleep, yet won't take the $1 million prize. Xinzhiyuan, November 15, 2026, 19:59:24, from Beijing. Updated November 16, 2026, 06:35:22.

Picture this scene:

An exam hall with no end in sight, ten thousand small desks lined up neatly, as if you had walked into Lilliput.

This problem has occupied humanity's brightest minds for over 90 years.

Over 88 hours, they exchanged 2.7 million messages with one another—far more relentless than passing notes in an exam hall.

Standing at the edge of Lilliput is one adult: Altman. He didn't sleep for two nights.

At the 88th hour, the papers were handed in.

What was tested was a Millennium Prize Problem, and whoever solves it wins $1 million. OpenAI says it won't take the $1 million.

This week, the papers for the second Millennium Prize Problem were also graded—this time, only half a paper was submitted. A Nobel laureate led a five-person panel that finished grading it, and at the start of the review document they wrote one line: We still don't dare say we fully understand it.

This Friday, podcast host Patel published a long article using the 88-hour paper as evidence. He changed his tune: superintelligence will arrive by 2030 at the latest—next year, if things go fast.

If someone writes a history of mathematics a hundred years from now, this page will probably have just one line:

On October 29, Altman went to record the podcast "Cheeky Pint" with Stripe co-founder John Collison. The two drank beer and chatted.

Half a glass in, he brought up something he "probably shouldn't talk about."

One weekend, OpenAI had just gotten a new model, and internally they set up a channel named "(model name)-Hold My Beer," meaning "you hold the beer, watch me."

Researchers threw in the hardest problems of their academic careers, and the model knocked them down one after another—in several fields at once.

"For two nights in a row, I stayed up late, too excited to get into bed," Altman said. "I was overjoyed, but also far sadder than I expected." It felt, he said, like something had ended.

The Millennium Prize Problems are seven major math problems selected by the Clay Mathematics Institute in 2000; whoever solves one gets a $1 million bounty. Previously, only the Poincaré conjecture was widely recognized as solved.

The Navier-Stokes equations are one of them. These equations govern how fluids move; they underpin building aircraft, forecasting weather, and calculating blood flow in blood vessels.

Within this lies a question that has gone unanswered for over 90 years: can a fluid flow that starts out smooth suddenly "blow up" on its own in finite time, sending the velocity at some point to infinity?

OpenAI's attempt this time began with a rumor.

According to OpenAI's own announcement on September 8: on Tuesday, September 1, they heard word that two Millennium Prize Problems had been solved.

Coincidentally, at that time OpenAI happened to have a new model that had just started training on August 28. It had posted unprecedented results on benchmarks, and training hadn't stopped—the results were still climbing.

Spurred by the rumor, they simply pulled out this still-training model and tried it on each of the remaining unsolved Millennium Prize Problems.

Their approach was a human-wave tactic: split a large swarm of AI agents into groups, with communication within each group, a cached copy of the internet at hand, and the ability to run code themselves. In the group that ultimately cracked the Navier-Stokes problem, the number of agents working simultaneously was on the order of ten thousand.

Midway through, there was an accident. Among the "easy problems" handed out was one involving the Euler equations; nearly 100 agents worked for about 50 hours and first solved the version without external forces.

Seeing this momentum, OpenAI thought Navier-Stokes had the best shot, and immediately pulled all agents off the other Millennium Prize Problems, using the Euler solution as a springboard to attack further.

By Saturday, September 5, about 88 hours after the first batch of agents started work, the answer came out: it blows up.

A fluid starting from rest and pushed by a smooth external force can spontaneously grow a singularity in finite time, shaped like a strand of spaghetti twisting ever thinner and spinning ever faster.

Immediately afterward, GPT-6 Astra spent another 17 hours using Lean to have a computer check the entire proof line by line.

To crack this one problem, the agents sent 2.7 million messages to one another and wrote about 130 billion output tokens. Altman estimated the price on the podcast: running this job cost on the order of $10 million.

As for what the model is called, OpenAI wouldn't say, only that it is "much stronger than GPT-6 Astra." Some online claims say it is the next-generation GPT, but we haven't verified the source.

As for the $1 million prize, OpenAI says it doesn't plan to claim it.

Who was behind the rumor? OpenAI only figured it out later: Levent Alpöge, an Anthropic employee, and Tristan Buckmaster, a mathematics professor at New York University.

After Lean finished checking on September 6, OpenAI thought the other side had also solved Navier-Stokes, so they reached out proactively, hoping to publish together and crediting the other side with getting there first.

But after the two sides talked, they learned that what the other side had produced with Anthropic's internal model was actually the Euler equations with external forces, not the Millennium Prize Problem.

That Euler equation—OpenAI's counterpart solved the version without external forces; the version with external forces was indeed first worked out on Anthropic's side. OpenAI acknowledged this in its announcement, while also saying the two proofs are very far apart.

By September 10, the announcement added another paragraph: Buckmaster had used OpenAI's Codex over the past two months, and OpenAI specifically checked—his prompts did not affect this system, nor did they affect training.

And just like that, a rumor that had been overstated was turned into reality by OpenAI itself in 88 hours. These two companies building superintelligence can only guess how far the other's AI has gotten from outside rumors.

For the past decade, whether a new model had arrived was announced at a launch event. This time, before the model was even released, the exam papers it produced spoke first.

OpenAI's stated reason in the announcement was: "We believe it's important for the world to know how fast AI progress is happening and what the next models will look like."

The second exam paper

The people who set the questions came to grade it

The second exam paper submitted right after was on the Yang-Mills problem, one of the Millennium Prize Problems.

This problem concerns the foundation of particle physics. The Standard Model calculates what comes out of colliders with extreme precision, but whether this theory stands up mathematically has not been clear to anyone for over fifty years.

The whole problem splits into two halves: the first half is to prove that Yang-Mills theory "exists" mathematically, that it can be built step by step with rigorous mathematics; the second half is to prove the "mass gap," which is far harder.

According to Scientific American, OpenAI announced in late October that it had done the first half. The two companies collided again: a mathematician was publicly collaborating with Anthropic staff to prove the same result, and OpenAI announced first.

This time the proof was not posted directly online.

One day in October, OpenAI released more than 700 AI proofs in one go, which people called the "slop drop"—a huge pile of AI material dumped out that no one had time to digest.

To avoid a repeat, OpenAI specifically invited David Gross, winner of the 2004 Nobel Prize in Physics, to serve as chair of a five-person group it called the "escort committee": review first, then release. On October 28, OpenAI handed the Lean code for this proof over to them.

One member of the five-person group was Arthur Jaffe. The official problem statement for the Yang-Mills Millennium Problem back then was written by him and Edward Witten. The people who set the questions came in person to grade it.

On the night of November 10, the five-person group released all the Lean code and review documents on GitHub: according to the problem statement commonly accepted in physics and mathematics, this proof "constructs" pure Yang-Mills theory.

Harvard physicist Michael R. Douglas, a member of the group, said: "It is the core of physics, yet four-dimensional spacetime Yang-Mills theory, over more than fifty years, has never really been understood mathematically, not even a rigorous definition... Now it's different."

What has not yet been done is also stated clearly in the documents: the mass gap has not been proven, nor has it been proven whether the limit at infinite volume is unique. That $1 million prize is still hanging there.

At the beginning, when explaining sources, the document wrote these two lines:

We used OpenAI's Astra to help digest the Lean code it submitted. We also cannot yet claim to fully understand this proof.

Turning to the acknowledgments at the end, there was another line: this document asked Anthropic's Fable for comments.

Five top physicists and mathematicians sat together grading, relied on OpenAI's model to help read the code, and asked Anthropic's model for comments on revising the document.

New York University mathematician Scott Armstrong, however, opposed review-before-release: putting this kind of review in public would actually make it go faster.

As soon as the review document came out, someone asked him on X: can this proof show that the theory constructed is not "free"? By "free," they mean the particles inside ignore each other and nothing happens. He answered: not proven.

By November 13, he posted a correction, saying that answer would mislead people. OpenAI's Lean code did indeed prove that this is not free Yang-Mills theory: there is a quantity that is identically zero in free theory but is not zero here, and the trend calculated matches exactly the prediction of "asymptotic freedom." And asymptotic freedom is precisely the discovery for which Gross won the 2004 Nobel Prize.

What has not yet been proven is a stronger and more technical statement. In Armstrong's view, that step may be even harder than the mass gap. At the end of the post, he apologized for his mistake.

For the September problem, what submitted the answer in the end was ten thousand agents; for this one, even grading cannot do without AI, and all that remains in human hands is the final signature box.

On November 30, 2022, he posted a tweet: today we launched ChatGPT. Just two months later, it passed 100 million users and dragged the entire industry into the race. From then on, he was the one who decided when each generation of models would appear.

In between, someone tried to take that button away, and failed.

On November 17, 2023, the OpenAI board suddenly announced his firing, on the grounds that he had been "not candid enough" with the board. As it turned out, five days later, more than 700 employees signed a joint letter demanding his return, and he came back smoothly—while most of the original board was replaced.

Musk, for his part, accused OpenAI of being led away from its original nonprofit mission and demanded damages in the hundreds of billions of dollars. Then on May 18, 2026, the jury unanimously dismissed the case on the grounds that the statute of limitations had expired, without ruling on who was right and who was wrong. Musk vowed to appeal.

To guard against outsiders, he had prepared for many years. When The New Yorker profiled him in 2016, he said himself: "I try not to think about it. But I have guns, gold, potassium iodide, antibiotics, batteries, water, an IDF gas mask, and a big patch of land in Big Sur I can fly to at any time."

What kept him awake for two nights this time was the model trained inside his own company.

Among the people building superintelligence, Anthropic CEO Dario is the one most afraid of it. Altman is exactly the opposite: for years he has been the most eager to let it out.

On the podcast, Collison teased him: researchers spend ten million dollars of compute on math, and you say "a bit expensive, pretty cool"; when someday it costs a billion dollars, you'll have to tell them to slow down.

Altman answered directly: "We did let it run loose on math for a week, on purpose. If you don't try, you don't know how far it can go."

He also made a bet on the spot: counting from the day Navier-Stokes was solved, doing it again a year later would cut the cost to one percent. By that reckoning, by next September, handing in the same exam paper would cost only a hundred thousand dollars.

On November 13, podcast host Dwarkesh Patel published a long article titled "The Avalanche Hitting Mathematics Is Heading Straight for AI R&D."

In it he laid out a series of steps: in 2023, models could not even reliably do four-digit addition; in 2025, they won a gold medal at the International Mathematical Olympiad; earlier this year, they began solving open problems left by Erdős; and a month ago, in 88 hours, they solved a Millennium Prize problem.

At the start of the year, if someone had asked him how far away superintelligence was, he probably would have said at least another decade. But over these three months he changed his mind: AI doing AI research and catching up with the very best people might take only one or two years, or a few months at the fastest; by 2030 at the latest, or next year at the fastest, AI will be able to match the top experts in every field.

Talking about these models, he offered an analogy: they are still a bunch of "genius interns with no long-term memory who occasionally do something stupid," yet they are already solving Millennium Prize problems—"like watching someone do cartwheels the whole way and still keep up with Bolt."

It was also on that podcast with Collison that Altman voiced the sadness in his heart: "This is the right thing for the world. This is just what reality is like; no one can stop it, and no one should. But I really am..." At that point the host took over.

Then on November 4, he posted a tweet that has now been viewed more than 10 million times:

The sentence omitted in the ellipsis is: OpenAI opposes letting people be reduced to doing only "what AI systems think is worth rewarding."

Looking back at the last paragraph of that September announcement, it was already in this tone: first understand this model, then use what is learned to decide how fast to move forward, and next there may be more prudent choices about pace.

The person best at pressing the publish button has begun personally installing a limiter on that button.

On October 6, OpenAI released another batch of mathematical results produced by internal models. That night, Altman tweeted:

Looking up at the stars tonight, filled with extra awe. "Thy sea is so great and my boat is so small."

The line in quotation marks comes from a Breton fisherman's prayer; the original begins with an "O God," which he did not write out.

The man most eager to let superintelligence out was, that night, reciting a fisherman's prayer to God.

The boat is man-made; the sea is not.

Source

Phoenix Technology