Artificial intelligence has been sold to businesses largely on a promise of efficiency: machines can perform intellectual work faster, cheaper and at a scale humans simply cannot match.
OpenAI’s latest mathematical breakthrough certainly demonstrated the faster and scale parts.
The cheaper part is a little more complicated.
In September 2026, OpenAI announced that an internal AI system had produced a solution to the Navier-Stokes existence and smoothness problem, one of mathematics’ famous seven Millennium Prize Problems. The Clay Mathematics Institute established the problems in 2000 and attached a $1 million prize to each successful solution.
According to OpenAI, approximately 10,000 AI agents worked concurrently on the Navier-Stokes effort. They reached the solution about 88 hours after the first agents were launched. During the Navier-Stokes work alone, those agents exchanged approximately 2.7 million messages and generated approximately 130 billion output tokens. OpenAI then used GPT-6 Astra for another 17 hours to formalize and verify the proof. (OpenAI)
That is an extraordinary technological achievement.
It is also an extraordinary bill.
The Guardian Puts the Cost at $15 Million
On September 12, The Guardian reported that OpenAI had “unleashed 10,000 agents” on the problem and that the bill was estimated at $15 million. (The Guardian)
OpenAI itself has not publicly disclosed an exact dollar figure for the project, so the $15 million should be treated as a reported estimate rather than an official OpenAI accounting.
Still, it gives us enough information to perform some entertaining arithmetic.
If 10,000 agents operated across an 88-hour effort, that represents a maximum normalized compute workforce of:
10,000 agents × 88 hours = 880,000 agent-hours
Put another way, that is the equivalent of one AI agent working continuously, 24 hours a day, for roughly 100 years.
Of course, 10,000 concurrent agents does not necessarily mean that every agent was consuming inference continuously for every second of those 88 hours. Agents can wait, communicate, execute tools, terminate and restart. So “880,000 agent-hours” is best understood as a useful normalization of the swarm, not a literal OpenAI billing unit.
But using that normalization and The Guardian’s $15 million estimate gives us:
$15,000,000 ÷ 880,000 agent-hours = approximately $17.05 per agent-hour.
So a reasonable shorthand for this particular experiment is:
One frontier AI research agent: about $17 per hour. Ten thousand of them: about $170,000 per hour. Run the swarm for 88 hours and you arrive at roughly $15 million.
Then There Are the Tokens
OpenAI says the Navier-Stokes portion of the experiment consumed approximately 130 billion output tokens. (OpenAI)
That’s 130,000,000,000 output tokens spread across approximately 880,000 normalized agent-hours, or about 148,000 output tokens per agent-hour.
If we simply divide the Guardian’s estimated $15 million total cost by the reported 130 billion output tokens, the all-in experiment works out to approximately $115 per million output tokens.
That should not be confused with OpenAI API pricing. The $15 million estimate presumably encompasses far more than just metered output tokens: input and context processing, model inference, orchestration among thousands of agents, infrastructure, tool execution and other computing resources.
But it gives us a sense of the computational scale required to get 10,000 artificial mathematicians arguing with one another until one group finally comes up with the answer.
About That $1 Million Prize
And here is where the story becomes particularly appropriate for the current economics of artificial intelligence.
Navier-Stokes is a $1 million prize problem. The estimated cost of solving it with AI? $15 million.
So, viewed strictly as a financial transaction: potential revenue $1 million, estimated expense $15 million, return on investment negative $14 million.
There is one small detail that makes the joke even better: OpenAI has said it doesn’t intend to claim the $1 million prize. (The Guardian)
So technically, OpenAI didn’t spend $15 million to win $1 million.
It may have spent $15 million solving a problem worth $1 million and then declined the $1 million.
That’s considerably more efficient.
And That Is AI Cost Effectiveness
Obviously, nobody at OpenAI ran this experiment because they needed the prize money.
The real payoff is demonstrating something potentially much more valuable: that massively parallel AI systems can attack research problems that have resisted generations of the world’s best mathematicians.
If this approach works repeatedly, the economics could eventually become transformative. Models will become cheaper. Hardware will become more efficient. Agent orchestration will improve. Instead of throwing 10,000 agents at every possible idea, future systems may get dramatically better at identifying the handful of promising paths.
But today, this experiment provides a wonderfully concise snapshot of the strange economics of frontier AI.
We have invented an intellectual worker that can operate for approximately $17 an hour. The catch is that sometimes you need 10,000 of them.
And so the AI industry has apparently discovered a revolutionary new definition of cost effectiveness: spend $15 million solving a $1 million problem, then don’t collect the million dollars.
Welcome to efficiency in the age of artificial intelligence.
