10,000 agents, 88 hours and a million-dollar problem.
AI gets serious about science

We have come quite a long way over the past few years.

At first, AI was only supposed to “put words together nicely”.

Then it turned out it could write code.

Later, it started running, testing and improving that code itself.

The next step was agents: instead of one answer to one prompt, a model gets tools, working memory and the ability to take successive steps, working independently on a task for longer periods.

In September 2026, we saw what may have been the first truly spectacular demonstration of what happens when you add one more ingredient:

scale.

OpenAI deployed around 10,000 AI agents working in parallel and set them to work on Navier–Stokes — one of the Clay Mathematics Institute's seven Millennium Prize Problems.

After approximately 88 hours the system had a solution.

This was followed by roughly another 17 hours of formalisation and verification in Lean, carried out with the help of GPT-6 Astra.

On 11 September, the Clay Mathematics Institute officially stated that the problem had “apparently been settled” — it appeared to have been resolved.

That does not mean the million-dollar prize has been awarded or verification completed. Clay's rules require publication, at least two years of public scrutiny and general acceptance by the global mathematical community.

But at this stage, this seems to be more than another case of “ChatGPT claims to have proved the Riemann hypothesis”.

Something much more interesting is happening.

What is the Navier–Stokes problem actually about?

I will not give a lecture on differential equations here. I am not a mathematician myself, and we do not particularly need one to understand why this story matters.

Very broadly, the Navier–Stokes equations describe the motion of fluids.

Water.

Air.

Flow around an aircraft wing.

Blood flow.

We have known the equations since the nineteenth century and use them practically all the time.

The problem is something else.

We do not know whether, in three dimensions, a solution starting from completely “normal”, smooth conditions can develop a mathematical singularity in finite time — a situation in which certain quantities grow without bound.

In 2000, Clay added this question to its list of seven Millennium Prize Problems and offered $1 million for solving each one.

Until now, one had been solved: the Poincaré conjecture.

For Navier–Stokes, the official statement permits several routes to a solution. One can prove global smoothness or construct a case in which a breakdown occurs.

OpenAI took the second route.

The system constructed a case in which a fluid starts at rest and experiences a smooth external force; its energy remains bounded, yet its velocity becomes unbounded in finite time.

This corresponds to variants C and D of Clay's official problem.

A small but important detail is worth stressing here: the familiar Navier–Stokes question without external forcing remains a separate open question.

The Millennium Prize Problem was deliberately formulated more broadly, however, and solving variant C or D is sufficient to settle it.

Astra did not do this

This is where the technology becomes interesting.

The solution was not generated by GPT-6 Astra, OpenAI's current public model.

The company explicitly says it used an internal model significantly more capable than GPT-6 Astra.

Its training began on 28 August and continued throughout the experiment.

Moreover, whenever a better-trained version became available during the work, OpenAI updated the running agents.

Astra entered the picture later, during formalisation and proof-checking in Lean.

So there are really two different tasks here.

First:

find a solution.

Then:

try to prove formally that the solution is actually correct.

The second is especially important with AI.

A language model can produce a very convincing-looking sequence of mathematical nonsense.

Lean works differently. It is a proof assistant: a system in which each step of a proof must follow formally specified rules.

That does not automatically close the matter. Among other things, we still need to check that the formalised theorem matches exactly what needs to be proved.

But it makes it much harder to hide a sentence on page 150 along the lines of:

“and now, of course, we assume X is true”.

10,000 agents are not one very clever chatbot

This may be the most important part of the story.

OpenAI did not type into ChatGPT:

“Solve the Navier–Stokes problem”.

The system consisted of thousands of agents.

They had access to a saved copy of the internet and could run code, among other tools. They were divided into groups that could communicate internally.

Different groups received different variants of the problem and were allowed to explore different approaches.

The best ideas were then collected and passed between groups. OpenAI calls this process cross-pollination — the cross-pollination of ideas.

Codex was among the tools used to consolidate the agents' results.

So instead of one exceptionally intelligent mathematician, we get something resembling a digital research institute.

One group tries method A.

Another tries B.

A third concludes that A leads nowhere.

A fourth spots an interesting part of B and combines it with something found by group seventeen.

The result then goes to the next hundred agents.

The whole process happens in parallel.

First, they solved an easier problem

This is even clearer in the way the system reached Navier–Stokes.

OpenAI also deployed agents in parallel on a set of problems considered easier.

One was a similar question for the Euler equations — broadly speaking, a relative of Navier–Stokes without viscosity.

Around 100 agents worked on it for roughly 50 hours.

They found a solution to the variant without external forcing.

Only then did OpenAI decide that this route looked most promising.

Resources were moved from other Millennium Prize Problems to Navier–Stokes, and the Euler result was given to the next agents as a starting point.

This essentially looks like ordinary scientific work.

You solve an easier case.

You extract a method from it.

You try applying it to a harder problem.

If something starts working, you increase the resources.

The difference is that a human makes a few such iterations.

Here, we can make thousands at once.

130 billion tokens

The scale of the experiment is almost absurd.

In the Navier–Stokes solving process alone, the agents exchanged around:

2.7 million messages

and generated around:

130 billion output tokens.

Across the entire experiment, including the other problems, the totals were 4.9 million messages and roughly 300 billion tokens.

The Verge, citing information from OpenAI, reports that the computing used for Navier–Stokes cost tens of millions of dollars.

That matters because it puts a little perspective on the claim:

“AI needed 88 hours for something humans had not solved in almost a century”.

Yes.

But those were a very particular 88 hours.

They involved around 10,000 agents and infrastructure to which practically no university research team has access.

Perhaps that is where we see the future more clearly than in the model itself.

We are no longer just scaling the model. We are scaling the thinking

Over the past few years, AI development has mainly meant increasingly large models.

More parameters.

More data.

More training.

Then came inference-time compute: a model could spend more computation on a harder question.

Now we are adding another dimension.

We are not asking one model to think for longer.

We are running thousands of copies of a model that try to think differently.

That is an enormous difference.

If I have a difficult programming problem, I can spend an hour searching for a solution.

Or I can hire ten good developers and tell each one:

“try something different”.

After an hour, we collect the results.

AI lets us take the same principle to an extreme.

Not ten.

Ten thousand.

And there is no need to arrange desks, meetings or coffee, or persuade group 8,347 to spend another three hours checking a hypothesis that probably makes no sense.

Then comes the controversy

Of course, the story has another layer.

OpenAI acknowledges that the experiment began on 1 September after it heard a rumourthat two Millennium Prize Problems might just have been solved.

One of those rumours concerned work by NYU's Tristan Buckmaster and Anthropic employee Levent Alpöge.

Both had been working on related problems for some time, also using AI.

Including OpenAI's Codex.

Buckmaster says their working materials were submitted to Codex.

Naturally, a question arose:

could OpenAI have used their work?

Following an internal investigation, the company makes a specific claim: Buckmaster's prompts from the two months preceding publication on 8 September could not have influenced the system in any way, including through training.

OpenAI explains that the model used in the experiment was produced through a large reinforcement-learning process run on a previously trained base model.

That sentence matters.

It tells us quite a lot about the last two months.

It does not tell us everything about the preceding years.

How many mathematicians had previously told AI where they were stuck?

Here we leave established facts behind and begin to speculate.

I want to make that very clear.

There is no evidence that earlier unpublished ideas from Buckmaster, Alpöge or other mathematicians were included in the model's training data.

Nor is there evidence that such data was what enabled AI to solve Navier–Stokes.

But there is another side to the story.

OpenAI itself says that in consumer services such as ChatGPT and Codex, user content may be used to train modelsunless the user opts out appropriately. Codex also has separate settings for training on the full environment.

Business products, Enterprise and the API have different rules: their data is not used for training by default.

Personally, I would be very surprised if mathematicians had not submitted thousands of conversations to models over the past few years along the lines of:

“check this proof for me”,

“where am I making a mistake?”,

“can this method be applied to Navier–Stokes?”,

“I'm stuck here — try to find another approach”.

Perhaps that included people now at the centre of this dispute.

Perhaps it did not.

We simply do not know from publicly available information.

But the underlying problem will only grow.

A friend over a beer and a million-dollar problem

A simple example illustrates it best for me.

Imagine I have been trying to solve a problem for twenty years.

Over a beer, I tell a friend:

“I'm really close now. I've reached these steps, but I keep failing at this one point”.

A few months later, my friend finds the solution.

Who solved the problem?

Me, because I did most of the work over twenty years?

Or my friend, because they found the missing step?

Even here, the answer is not obvious.

Now let us make the experiment more complicated.

It is not one mathematician talking to one friend.

For five years, hundreds or thousands of mathematicians tell AI over a beer where they are stuck.

Some do it publicly in articles, forums and papers.

Some may do it in conversations with models.

One shows idea A.

Another explains why A does not work.

A third finds variant B.

A fourth gets almost to the end of C.

Then a model is trained on an enormous amount of mathematical knowledge.

We run 10,000 copies of it.

And say:

“you have 88 hours”.

If the system finds a combination of ideas no human previously found, who is the author of the discovery?

I do not have a good answer.

And I suspect our current rules on authorship do not have one either.

A checkbox starts to have real value

There is another very practical lesson in all this.

Privacy settings in an AI tool are becoming more than cosmetic.

If you submit to an external model:

unpublished research,

product code,

company know-how,

technical documentation,

a patent,

an algorithm idea,

it pays to know what happens to that data.

Is it used for training?

Or not?

Are you using a consumer version?

Or a business version?

How is Codex configured?

Not because OpenAI will steal your start-up tomorrow.

Because with today's model capabilities, information really can have value.

In an ordinary conversation, the difference between a checked and unchecked box may have no practical significance.

For research potentially worth a million dollars, it is rather different.

The controversy is interesting. The technology is more interesting

The dispute over Buckmaster, Alpöge and OpenAI will probably continue for some time.

There are questions about authorship.

Access to data.

How conversations are handled.

Whether a company supplying tools to scientists can later become their competitor.

The Verge spoke to more than a dozen mathematicians, and a recurring concern is something we might call the industrialisation of scooping.

Previously, the information:

“team X is close to solving the problem”

did not give competitors much of an advantage.

They still had to find a few people with the right expertise and spend months or years working on the same method.

Now, potentially, you can hear a rumour on Tuesday and launch several thousand agents.

That completely changes the dynamics of research.

But despite the controversy, the technical part of this story seems far more important to me.

This is not an artificial Einstein yet

Of course, it is easy to get carried away in the other direction.

“AI solves problems in three days that humans could not solve in a hundred years”.

Not quite.

The system did not start from scratch.

It had access to a huge portion of existing mathematical knowledge.

Humans chose the problem.

Humans designed the architecture of the whole experiment.

The agents used tools and existing results.

Enormous amounts of computing power were consumed.

And the result still needs scrutiny from the mathematical community.

So we have not seen a lone digital Einstein gazing out of a window, going “hmmm” and inventing new mathematics.

We may have seen something more practical.

A research factory.

A system capable of running thousands of parallel attempts, letting them explore different paths, collecting the best ideas, passing them on and finally using formal tools to check the result.

That could be a much more important technology than one exceptionally intelligent chatbot.

88 hours that could change more than one problem

OpenAI already says that after finishing its Navier–Stokes work, it made “substantial progress on another Millennium Prize problem”.

We do not yet know which one.

Perhaps nothing will come of it.

Perhaps another paper will arrive in a few days.

There is no point guessing.

But if the way of working OpenAI describes proves repeatable, the consequences will extend well beyond mathematics.

A similar model is imaginable wherever we can:

generate thousands of hypotheses,

reject incorrect ones relatively quickly,

run simulations or code,

compare results,

and pass the best ideas to the next agents.

Mathematics is an almost ideal testing ground in this respect, because at the end we can ask:

is the proof correct?

In biology, chemistry or materials science, verification may require a physical experiment.

But the search mechanism remains similar.

That is why I do not think the most important question in this story is:

“has AI just won a million dollars?”

OpenAI, incidentally, says it does not intend to apply for the prize.

A much more interesting question is:

what happens when 10,000 agents like these stop being a one-off experiment costing tens of millions of dollars and become an ordinary research tool?

Then AI will no longer just help scientists.

It will become another participant in the scientific process.

Alongside it, we will have to revisit some very old questions:

Who made the discovery?

Who deserves authorship?

Whose idea was it?

And what does “my work” actually mean if I spent years discussing it with a model beforehand?

Those 10,000 agents do not seem to have answered these questions for us yet.