What did OpenAI claim about Navier–Stokes?
On 8 September 2026, OpenAI published a post titled "On the Navier–Stokes Millennium Prize Problem". The company said an internal system had produced a proof that the equations describing fluid motion can break down in finite time.
The claim is specific. OpenAI says an initially smooth fluid at rest, with a smooth force applied and finite energy throughout, can develop a singularity. In its words, the result establishes statements "C" and "D" in the official Millennium Prize formulation.
The mechanism is a vortex that spirals inward and stretches, like spaghetti. The fluid speeds up without bound while its energy stays finite. OpenAI argues the blow-up comes from the motion itself, not from an infinite force inserted by hand.
The timeline is unusually short. OpenAI says the effort began on 1 September, after it heard rumours that two Millennium Prize problems had been solved. Its agents reached the result on 5 September, about 88 hours after launch.
Lean formalisation and verification took a further 17 hours with GPT-6 Astra. The group behind the result involved "on the order of 10,000 concurrent agents". The Navier–Stokes run alone sent 2.7 million messages and consumed about 130 billion output tokens.
Across every problem attempted, the agents sent 4.9 million messages and used roughly 300 billion output tokens. TechCrunch estimated that output would cost about $22.5 million at current Astra rates, though OpenAI did not bill it that way.
There was a warm-up act. The system first resolved the unforced version of the Euler regularity problem, a related question reached by removing viscosity. OpenAI says nearly 100 agents worked for about 50 hours on that result, which persuaded the team to redirect resources to Navier–Stokes.
OpenAI also said the model behind the result is "significantly more capable than GPT-6 Astra", and that its training began on 28 August. The company added that it does not intend to claim the Millennium Prize.
Why do mathematicians dispute the announcement?
Hours before OpenAI's post went live, Tristan Buckmaster, a mathematician at New York University, published his own statement. He had been working with Levent Alpöge, a mathematician employed by Anthropic, on a related question involving the Euler equations.
Their drafts sat in OpenAI's Codex coding tool, which they used to develop the work. Buckmaster wrote that he did not know what OpenAI's model did, or whether his data was used. He also wrote a line that has travelled far: "I am not accusing anyone of anything."
His suspicion concerns the route, not only the timing. The pair had attacked the problem through a smooth force, options C and D in the official statement. Buckmaster said of that choice: "Almost nobody else I know of was working on it."
OpenAI's account confirms part of the timeline. It says it began on 1 September after hearing a rumour about the pair, and that it contacted them after finishing its own project and verification on 6 September. It offered a concurrent release and recognised their priority on the forced Euler result.
The company says nobody at OpenAI, human or AI, saw the other team's work before publication. It also says no specific user data was accessed for the proof.
That answer shifted. In its first version, OpenAI wrote that while contamination was unlikely, it could not rule out that de-identified data from the pair's use of its products helped improve its models. On 10 September it updated the post. An investigation had confirmed, it said, that Buckmaster's Codex prompts from the previous two months could not have influenced the system, including through training.
Buckmaster's statement also describes private exchanges. He says OpenAI researcher Sébastien Bubeck asked him to drop Alpöge's credit as part of a compromise. According to his account, Bubeck then asked: "Why would you ruin your career?" OpenAI's public post does not address that claim.
Bubeck, at a press briefing reported by The Guardian, called the result "a spectacular culmination of the arc we have seen over the past 12 months".
Some of the disagreement is technical rather than personal. Alpöge and Buckmaster produced a result for the forced Euler equations. OpenAI's system produced one for the unforced case, plus the Navier–Stokes proof. OpenAI says the proofs differ significantly, and that the precise results are different.
The unresolved question is not whether two teams worked at once. It is whether prompts stored on a company's servers can reach the model that company then deploys. Neither side has published logs that would settle it.
What is the Clay Institute doing?
The Clay Mathematics Institute named Navier–Stokes existence and smoothness one of seven Millennium Prize Problems in 2000. Each carries a $1 million prize. The problems were unveiled in Paris to mark the turn of the millennium.
Clay responded on 11 September. It said it shared "in the excitement of the global mathematical community" at an announcement that the problem had "apparently been settled". It added that the process for evaluating credit is "deliberately unhurried", and promised updates.
The rules explain the delay. Before Clay will consider a solution, three conditions must be met.
That means no prize can be awarded before 2028, even if the proof survives scrutiny. As of 28 October, Clay's news page still shows the 11 September statement as its most recent item. No verdict has been published.
Formal verification does not settle the question by itself. A Lean proof shows that a theorem follows from its formal statement. It does not show that the formal statement captures the problem mathematicians meant to ask. That gap is one reason Clay relies on refereed publication and community acceptance rather than a machine check alone.
- •The solution must be published in a qualifying outlet, as defined in Clay's rules.
- •At least two years must have passed since that publication.
- •The solution must have received general acceptance in the global mathematics community.
How is AI changing scientific publishing?
The BBC's investigation, published on 28 October, widens the frame from one proof to the whole research pipeline. It reports that journals are struggling to screen a wave of AI-generated text, images and references.
The strongest evidence comes from biomedical abstracts. Dmitry Kobak, a computer scientist at Ghent University, and colleagues studied more than 15 million PubMed abstracts from 2010 to 2024. Their analysis, published in Science Advances, found that at least 13.5% of 2024 abstracts had been processed with a large language model. For some subsets the lower bound reached 40%.
The BBC reports a higher figure from Kobak's team. By December 2025, it says, nearly 90% of biomedical papers showed AI fingerprints. Detection is imperfect, and the BBC notes that some tools have labelled the US Constitution, written in the 1780s, as AI-generated.
Kobak told Nature that the risk is systemic rather than individual. "Whatever bias the LLM may have will just suddenly permeate the (scientific) literature," he said.
Peer review is affected too. A Stanford team led by Weixin Liang and James Zou analysed reviews submitted to four major AI conferences: ICLR 2024, NeurIPS 2023, CoRL 2023 and EMNLP 2023. They estimated that between 6.5% and 16.9% of review text could have been substantially modified by a language model.
Citations are the third weak point. The BBC reports that a Cornell University team found around 147,000 phantom works cited last year, invented by AI tools used to compile reference lists. A fabricated reference can survive in a paper for years before anyone checks it.
Publishers are responding with screening rather than bans. Elsevier has built about 100 signals that flag a paper for closer human review, according to the BBC. Sarah Main, an Elsevier vice-president, described the goal as keeping the researcher in control. "What we're calling research-grade AI would focus on ensuring the researcher is in the driving seat," she said.
Some rules are blunter. Nature's journals forbid AI-generated images and tell referees not to upload manuscripts into AI systems. arXiv bans researchers for a year if they submit invented references.
Dalmeet Singh Chawla, a journalist who reports on research misconduct, told the BBC that generative AI has "basically disrupted the whole scientific process". Asked whether publishers' integrity teams are keeping pace, he said: "Right now, it seems like they're not."
Does AI produce reliable science?
The record is mixed, and the strongest example is still protein folding. Google DeepMind's AlphaFold, first presented in 2020, predicts the shapes of proteins, which determine what they do. DeepMind says the model predicted 200 million protein structures, work the BBC says would have taken hundreds of millions of years in laboratories.
The results were free to researchers. Demis Hassabis, DeepMind's chief executive, and John Jumper, the project's lead scientist, shared the 2024 Nobel Prize in Chemistry. Pushmeet Kohli, who leads science work at DeepMind, called AlphaFold a "watershed moment".
Materials came next, with more contest. DeepMind's GNoME model predicted 2.2 million new crystals, a body of work the company described as "nearly 800 years' worth of knowledge". Of those, 380,000 were rated stable enough to be worth making.
Two chemists at the University of California, Santa Barbara checked the claim. Anthony Cheetham and Ram Seshadri sampled the predicted compounds and published their assessment in Chemistry of Materials. They found "scant evidence for compounds that fulfill the trifecta of novelty, credibility, and utility", and described many as "trivial adaptations of known materials".
DeepMind rejects that conclusion. It says it tested every compound for novelty, and that dismissing a structure for looking familiar misses how materials improve. Swapping a single atom, the company argues, is how many useful materials were found, from phone components to sunscreen.
The disagreement is instructive. Both sides accept the method. They differ on whether the output counts as discovery before a human has made and measured the material. The published count is a prediction, not a shelf of new substances.
There are cases where AI helped answer a long-standing question. Andrea Pauli at the Vienna BioCenter spent almost ten years on how a sperm enters an egg. Her lab found a protein on the surface of zebrafish eggs that points to a lock-and-key mechanism, with AI assisting the search.
Kohli's own test is about accountability, not capability. "The agents today are imperfect. Somebody has to take accountability," he told the BBC. He compared abandoning AI to doing computation without calculators, and said researchers must understand what an agent did and take responsibility for it.
Who is affected, and what are the limits?
Mathematicians are the first group. Priority decides credit, jobs and funding, and a proof announced by a company can reset a career's worth of work in a week. Buckmaster's statement is as much about attribution as about mathematics.
Publishers and editors carry the cost of screening. Integrity teams must check text, images and references at a scale manual review cannot match. Every false positive also costs a researcher time, and the BBC notes that detection tools can misclassify human writing.
Funders and the public absorb the waste. The BBC points out that phantom citations send researchers down blind alleys, and that public money pays for the detours. Drug discovery and green technology move more slowly when the literature cannot be trusted.
Patients and students are affected indirectly. Basecamp Research, the London lab the BBC visited, is collecting DNA to build models for biology. Its co-founder, Oliver Vince, said the company has "a technically feasible path to curing cancer". That is a direction, not a result.
The limits start with access. Outside researchers cannot inspect the training data or the internal model, so claims about contamination rest on company investigations. In this case OpenAI reviewed its own system, and the public has the summary rather than the logs.
Detection is probabilistic. A flag for AI writing is not proof of misconduct, and the tools misfire on old human text. Formal proofs remove arithmetic errors, not errors of interpretation, and a verified Lean file still depends on a human-written statement.
Resources are uneven. TechCrunch's $22.5 million estimate for one week of compute shows why a small academic team cannot compete on raw scale. That asymmetry shapes which results get announced first, and who is left to verify them.
None of this makes AI useless in science. It makes verification the scarce resource, and the institutions that provide it are the ones now under strain.
What to watch next
The Clay review is the clearest test. Watch for refereed publication, independent checking of the Lean formalisation, and signs of general acceptance. Under the rules, the earliest possible award date is two years after publication.
Watch how OpenAI documents the model. The company says the system is more capable than GPT-6 Astra and that training is ongoing. Whether it publishes evaluation details, or only a proof, will shape how much other researchers can check.
Watch publisher tooling. Elsevier's 100 signals, Nature's image rules and arXiv's one-year bans are early controls. Their error rates, and whether they are published, will decide whether they build trust or just add friction.
Watch for independent audits of training data. The dispute turned on whether prompts stored in Codex could reach a model. A verifiable answer would need logs and access that companies have not offered.
Watch the science itself. If the Navier–Stokes proof holds, it changes fluid dynamics. If it fails, the episode becomes a case study in how quickly a company claim can outrun verification.
The BBC's broader question is harder to settle. AI can accelerate discovery and also flood the record with confident errors. The institutions that separate the two are the same ones now under strain.