Ten thousand AI agents exchanged 2.7 million messages over 88 hours, burned $15 million in compute, and generated roughly 130 billion output tokens. The product was a 165-page document that OpenAI says solves the Navier-Stokes existence and smoothness problem — a $1 million Millennium Prize question open since Jean Leray's work in 1934.
That was one of two documents OpenAI put into the world in the same window.
The other was a report on recursive self-improvement — how much its AI is speeding up its own research. It is far less quotable and far more honest. Read together, the two tell opposite stories about what AI can actually do in science.
Read together, the two documents tell opposite stories: one where AI cracks a 92-year-old problem, one where it buys you a modest 10 percent and can't be left alone.
Start with the fine print the announcement hoped you'd skip.
The problem "is not yet officially resolved," as even the triumphant coverage concedes. The 165-page proof has not cleared independent examination, let alone the Clay Mathematics Institute's process. OpenAI says it won't claim the prize money — a curious hedge for a company declaring victory.
Then there is the provenance question. NYU mathematician Tristan Buckmaster has accused OpenAI of throwing massive resources at the problem only after learning that he and Anthropic researcher Levent Alpöge were making progress on a related approach.
OpenAI has acknowledged the project began after it heard rumors that outside mathematicians had made progress. It denies accessing their unpublished work — while conceding it "could not rule out" that anonymized data from those researchers' use of OpenAI products had helped improve its models.
That is not a small caveat. It is the difference between an AI that discovers and an AI that overhears.
Diego Cordoba, whose work helped pioneer the approach the AI used, said a valid solution would be a "big surprise." When the person whose method underpins the proof calls a success surprising, that is a signal worth more than a press release.
A separate objection has reportedly been raised, in a post attributed to Nature's account: that the claimed proof likely constructs a geometric anomaly the laws of physics don't allow — a singularity of the kind normally hidden behind an event horizon. We won't adjudicate that fluid-dynamics question here. It matters for a different reason: it is exactly the kind of subtle failure that only survives contact with expert review, not with a compute budget.
Because the tell is in the method. Critics argue OpenAI tackled the problem largely by brute force — thousands of agents searching mathematical ground human researchers had already mapped, rather than through genuine reasoning or scientific intuition.
Ten thousand agents. 2.7 million messages. $15 million. 130 billion tokens. Those are inputs. None of them is a proof. A correct 165-page argument and an elaborate 165-page mistake consume roughly the same compute.
Which is precisely the problem with judging science by its energy bill.
Now turn to the second document — the one where the honest arithmetic lives.
OpenAI's recursive self-improvement report disclosed that the number of experiments per researcher in August 2026 was 1.6x what it was in 2025. That is the cleanest single finding in the whole release.
A 1.6x jump in experiment velocity sounds like acceleration. But R&D output doesn't scale linearly with experiment count. A large body of literature — inside AI and in innovation broadly — finds power-law, diminishing returns.
Run that 1.6x through a plausible power law, and you get, reportedly per Ramez Naam's own speculative assumption, roughly a 10 percent acceleration in actual AI progress. Not 60 percent. Ten.
The report's second chart measures success rate by task length. At every task length, some human intervention is sometimes required. Even on very short tasks, the AI isn't consistently, fully autonomous. The longer the task, the more often a human has to step in.
The third chart is the most revealing. It breaks down what kind of work the AI actually produced, by output tokens.
The AI did lots of code writing, lots of debugging, lots of launching runs, lots of assistance. It did much less deciding what to work on, planning experiments, or judging what to continue and what to stop.
In other words: the machine writes the code and runs the jobs. The humans still choose the questions.
The analyst Ramez Naam, reading the report, put the conclusion plainly: AI is "still mostly a helper, not a driver, and that any acceleration loop is, as of now, still pretty weak."
Hold those two documents side by side. In one, AI is the protagonist that cracked a 92-year-old problem in 88 hours. In the other, AI is the tireless research assistant that buys you 10 percent and can't be left alone on a long task.
The RSI report is the control experiment for the Navier-Stokes claim. And it quietly contradicts it.
Because if AI still can't reliably decide what to work on inside OpenAI's own labs — where it has every advantage — the story of an autonomous system independently conquering a Millennium Problem strains against the company's own measured data.
Notice which document got the headline and which got a thread. The jubilant, unverified claim traveled the world. The sober, data-heavy report circulated among specialists.
That asymmetry is the story. And the timing sharpens it: critics including prominent commentator Mehdi allege the announcement was engineered as a "Sputnik moment" to inflate excitement — and valuation — ahead of a potential stock-market listing.
Whether or not that motive holds, the incentive structure is real. A firm approaching an IPO has every reason to announce a breakthrough and no obligation to wait for peer review.
Which brings us to the actual lesson, and it has nothing to do with fluids.
Scientific credibility is a verification regime, not a press release. The Poincaré Conjecture — the only Millennium Problem resolved so far, by Grigori Perelman — reportedly became "solved" not when a claim was posted but after years of expert scrutiny confirmed it.
A "computer-verified proof," OpenAI's phrase, verifies that steps follow formal rules. It does not verify that the theorem states what the prize requires, that the construction is physically admissible, or that the approach wasn't quietly inherited from humans who got there first. Those are exactly the judgments the RSI report shows AI still leaves to people.
So here is the reasonable posture. When the same company publishes both a triumphant unverified claim and honest internal data, trust the data.
The data says AI is a powerful assistant with a weak acceleration loop and a reliability problem on long tasks. That is genuinely important — and genuinely less thrilling than "AI solves Navier-Stokes."
Until the proof clears independent peer review, no one should say a Millennium Problem has been solved. Not the company, not the coverage, not us.
Ten thousand agents can generate 130 billion tokens in 88 hours. They cannot confer certainty. That still takes the slow, unglamorous machinery of people checking each other's work — the one part of science no token count has yet replaced.