On 8 September OpenAI said 10,000 of its agents, running for 88 hours, produced a 166-page proof for the Navier-Stokes equations, a problem open since 1934 and one of the seven Millennium Prize Problems. The Clay Mathematics Institute, which set those problems in 2000, now says the problem appears to be solved while formal verification continues, and has marked it "active" on its site: no longer unsolved, not yet solved. The institute describes its own process as "deliberately unhurried." The distance between 88 hours of generation and a verification schedule measured in years is the part of this story that reaches companies that have never touched a partial differential equation.
The prize terms explain the pace. Each Millennium Problem carries $1 million, and the money has been awarded once. To qualify, a proof must first appear in a peer-reviewed journal of international standing, then survive two years of mathematicians trying to break it. Those conditions were written in 2000, and they do not ask whether the argument is formally correct. They ask whether mathematicians read it, discussed it, and failed to refute it.
OpenAI said it will not claim the money. That is an inexpensive concession. Outside estimates put the compute bill for the run between $10 million and $40 million, ten to forty times the prize. The gesture reads less like modesty than like an acknowledgment that the $1 million was never what the run was for.
On 11 September, 25 Fields Medal laureates signed a statement saying that the way AI labs approach mathematical problems is damaging mathematics. The following day Terence Tao, one of the signatories, handed his blog to two philosophers of mathematics, Silvia De Toffoli and Eamon Duede, who set out what the statement's authors meant.
Their argument turns on two things mathematicians rarely had to separate. A logical proof is deductively correct reasoning that a machine can verify. A comprehensible proof is reasoning from which people understand why a statement is true, can explain it, and can build further work on it. Historically nobody could produce the first without already holding the second, because constructing the argument required the understanding. AI has split them. On the philosophers' reading, OpenAI's result supplies an answer, and whether that answer has become a substantive solution anyone can work from is still open.
The same split is now visible inside companies with no interest in fluid dynamics, and there the numbers are less abstract.
Digital Applied's first-quarter 2026 survey covered 2,847 developers who actively use AI coding tools. They spent an average of 11.4 hours a week checking AI-generated code and 9.8 hours writing new code. Review is the larger job. LinearB, analyzing 8.1 million merge requests, found that agent-authored requests waited 5.2 times longer for a first review than requests with no AI involvement.
Workday's January survey of 3,200 employees and business leaders found the same shape outside engineering. 85% of employees said AI saves them between one and seven hours a week. 37% of that saved time went back into fixing, rewriting and checking the AI's output. Only 14% said they consistently see a clear positive net effect. And 77% of daily users check AI-produced work at least as carefully as a colleague's.
The work did not disappear. It moved from production, which companies measure, into verification, which they mostly do not. The output shows up in a productivity report; the effort needed to make it trustworthy dissolves into the working day.
Explainability is the word companies reach for here, and it answers a different question. Explainability asks whether a result can be traced, interpreted or justified. Comprehension asks whether the people accountable for the decision understand the result well enough to challenge it, defend it and extend it. An organization can satisfy the first completely and fail the second completely.
Levent Alpöge and Tristan Buckmaster, who published their own AI-assisted proofs the same week, called the first write-up of one of them the worst exposition in the history of mathematics. The Lean proof assistant had confirmed every line of it. Making the reasoning readable took weeks, and the text was still being worked on at publication. That translation step, from machine output into something a person can hold in their head, is where the time goes. In academia the labor has a name and a career attached. In business, the most expensive people on the payroll spend their days on it instead of making anything new.
I think the epistemology framing, which is where this debate has settled, is doing companies a favor they have not earned. The verification gap is not mainly a puzzle about the nature of proof. It is a staffing fact with a straightforward shape: generation scales with compute and judgement does not. The people who can catch a confidently stated error are the ones with enough domain knowledge to know what the answer should look like before they read it, and they are frequently the ones a company has in mind when it books the efficiency gain. Cutting them is not a lagging consequence of adopting AI. It is the move that converts a productivity story into a liability.
Notably absent from OpenAI's announcement is any figure for the other side of the ledger. The company counted agents, hours and pages. It did not say how many hours of human mathematical attention those 166 pages have absorbed since 8 September, or how many they will need before the Clay conditions are satisfied. That is the one number that would tell you whether 88 hours was the cost or the deposit.
Founders now get some version of the same question every week: what does a person have to do after the model answers, before a customer can use the result? Few can name a figure. Three habits would produce one. Count reading the way you count writing, so that hours spent checking, rewriting and approving AI output sit in the same ledger as the code, reports and closed tickets it generates; without that line, volume produced measures throughput rather than productivity. In review, ask who here understands this well enough to challenge it, defend it and extend it, and treat "nobody" as the answer it is: the company has an artifact, not a capability. And treat reviewers as a scarce input rather than an overhead, because they are the constraint on everything downstream.
Mathematics spent centuries building the machinery that separates an answer from accepted knowledge: journals, referees, public criticism, and time. The Clay conditions are that machinery written down, and they are about to be applied to a document produced in 88 hours by a system that will produce the next one faster. Most companies skipped the machinery and built a dashboard for generation speed, which is why the year is gone and nobody can say where.