AI and mathematics: what just happened, and what is left for humans?

In the first week of September 2026, an AI system built by OpenAI produced a proof about the motion of fluids in roughly 88 hours of agent work. Four weeks later, on 6 October, the company posted 722 more mathematical manuscripts in one go, grouped into 372 families of results. Mathematicians quoted by The New York Times, as relayed by TorontoToday on 17 October, called the release breathtaking and devastating.

Meet Claude-Louis Navier

The scale is real. So is the fact that a large part of it has not yet been checked by a human referee.
Claude-Louis Navier began writing down the equations of fluid motion in 1822, and George Gabriel Stokes completed the work by 1850. In the two centuries since, nobody has been able to prove that their solutions always stay well behaved, which is why the Clay Mathematics Institute made the question one of its seven Millennium Prize Problems in 2000.

Its own description of the problem says human understanding of these equations remains minimal.

So now what about prime numbers?

The targets included the Riemann hypothesis, which Bernhard Riemann posed in an 1859 paper on prime numbers and which is still widely called the most important open problem in pure mathematics, and a question descended from the fifth problem on David Hilbert’s 1900 list. Riemann’s hypothesis concerns the distribution of the prime numbers, which is about as old as mathematical questions get: Euclid proved in the Elements, around 300 BC, that the primes never run out. Mathematicians quoted by The New York Times, as relayed by TorontoToday on 17 October, called the release breathtaking and devastating.

The scale is real. So is the fact that a large part of it has not yet been checked by a human referee.

This piece sorts what is established from what is claimed in the story of AI and mathematics, and says what it means the next time you read the words “AI solved it”. Facts are as of 24 October 2026.

What happened in details

OpenAI’s results arrived in two steps: one headline proof in September, then a bulk release in October.

From one proof to 372 families

On 8 September 2026, OpenAI said an internal model had resolved the Navier–Stokes existence and smoothness problem, one of the seven Millennium Prize Problems named by the Clay Mathematics Institute in 2000. By the company’s account, the proof shows that a smooth fluid at rest, pushed by a smooth force, can develop a singularity: speeds that grow without bound in finite time. OpenAI says this establishes statements C and D in the official problem formulation. It published the paper and a Lean formalisation, and said it does not intend to claim the $1 million prize.

On 6 October the company posted 722 manuscripts to a public GitHub repository, grouped into 372 families of results. The topics span number theory, algebra, topology, logic and theoretical computer science. Reporting on the release lists claimed partial results on the Riemann hypothesis, the Birch and Swinnerton-Dyer conjecture and the Hodge conjecture, and an independent catalogue of the release lists a claimed solution of the Unique Games Conjecture.

The repository itself says the collection holds results at different stages of verification.

How fast was fast

The speed is what unsettles people, so the numbers matter. According to OpenAI, the Navier–Stokes effort began on 1 September, the agents reached a result on 5 September after about 88 hours, and formalising it in Lean took another 17 hours. The group that found it involved on the order of 10,000 agents running at once and about 130 billion output tokens.

For the October release, OpenAI says each result used on average the equivalent of three hours of ChatGPT Pro thinking time, across roughly 4,000 problems put to an unreleased model. The 372 families are what remained after the company required what it calls an appropriate level of significance.

So “solved in days” is accurate for the September proof. For the October pile it holds only per result, and only if you leave out the attempts that produced nothing, and the training of the model itself, which began on 28 August.

What is confirmed and what is only claimed

Only part of the release has been machine-checked, and the first errors surfaced within a day.

What the Lean check covers

Lean is software that checks a proof step by step against fixed rules, so a proof it accepts cannot hide a logical gap. OpenAI says many but not all of the manuscripts come with Lean files, and a company spokesperson told Retraction Watch that roughly half of the results were released without that confirmation. The repository’s own notes say some unformalised results could contain issues.

Lean also has a limit that is easy to miss. It certifies that a proof is valid for the statement that was typed in. A human still has to confirm that the typed statement says what the headline says. For Navier–Stokes, that reading is now the work of the mathematical community, and the Clay Mathematics Institute had not recognised the result in the reports we found from September.

Three withdrawals in a day

On 7 October OpenAI withdrew three manuscripts after finding a sign error that broke one argument and a construction used by two dependent papers. It also revised 14 others, Retraction Watch reported on 8 October.

Alex Townsend of Cornell told Retraction Watch he expects more errors to appear, and that the Lean-checked papers should have been announced first, with the rest offered as a request for help. Andrew Sutherland of MIT called the quick withdrawal the responsible thing to do, and added that regaining trust will take much more than that.

OpenAI says the errors were found in an audit and that it will withdraw any paper it cannot repair. The model has not been released, so nobody outside the company can re-run the work. The company says it is working to release it responsibly.

The human side

Mathematicians are not one voice on this, and the split does not run between friends and enemies of AI.

A rainforest with a hotel next door

The computer scientist Scott Aaronson, of the University of Texas at Austin, compared the moment to a lifelong forager who knows every path of a rainforest and then finds a resort hotel with a helipad and heated pools built beside him. Melanie Wood, a member of OpenAI’s advisory group, described the release as the start of the process of human understanding, not its end. Tristan Buckmaster of NYU said the company had not done proper due diligence, given how much it released at once.

On 7 October the Association for Human Mathematics went further. Its statement said mathematicians did not ask for this work, and urged them to stop working with OpenAI, according to Retraction Watch.

Tao and “Math 2.0”

Terence Tao, a Fields medallist who has worked openly with AI tools, has become one of the sharpest critics of how this was done. As reported by Fortune, his argument is that harvesting open problems at industrial scale leaves whole areas less fertile than if people had solved them slowly. He calls for a “Math 2.0” that rewards explanation and community, not only the count of solved problems. He has also said the unannounced release disrupted some junior researchers’ projects, according to EconoTimes.

In September, a group of Fields medallists published a declaration titled “A Severe Misalignment of AI in Mathematics”, warning against treating unsolved problems as benchmarks. Retraction Watch reports that more than 8,000 researchers have endorsed it. The objection is to the method and the tempo, not to the tool.

Who got there first

The September announcement came with a priority dispute. Buckmaster and Levent Alpöge, who works at Anthropic, had resolved a related question, the forced Euler equations, with an internal Anthropic model. OpenAI says its agents solved the unforced Euler problem and then Navier–Stokes, that none of its researchers or agents saw the other team’s work, and that an investigation found Buckmaster’s Codex prompts could not have influenced its model.

Buckmaster has alleged otherwise. These are two accounts, and the evidence that would settle it, the model and its training data, is not public.

What comes next

October held two stories at once: a bulk release that many mathematicians did not ask for, and a proof that mathematicians steered themselves.

The KLS conjecture, with people in the loop

On 21 October 2026 Microsoft Research reported that Xinzhi Zhang and Zhao Song had proved the Kannan–Lovász–Simonovits conjecture, posed in 1995. It asks whether, in any number of dimensions, the best flat cut through a broad class of shapes is nearly as good as the best cut of any kind.

By Microsoft’s account, AI agents explored more than 100 candidate approaches while the researchers pruned and redirected the search. Separate, unguided attempts by leading models with a similar compute budget fell short. Microsoft says independent experts posted their own proofs the same day, and a third-party Lean formalisation followed. This is the company’s own account, but the independent proofs are public on arXiv and open to checking.

Both stories use the same technology. In one, a model works alone for about three hours per result and posts the output. In the other, people decide where to dig and a machine does the digging.

Where that leaves human mathematics

What looks scarce now is not the answer but the surrounding work: choosing which questions matter, explaining a result so that others can use it, and checking it. On 15 October OpenAI said it would contribute $5 million to a Simons Foundation initiative, part of a $15 million effort, to help mathematicians examine and build on AI-produced results through workshops, conferences and special programmes, with access to OpenAI’s models and compute for the researchers it supports.

A mathematician interviewed by TorontoToday on 17 October offered a limit to the story: the problems involved are prestigious but not urgent, and AI cannot currently propose sensible solutions to urgent ones such as cancer or climate change.

Nobody can yet say how many of the 372 results will hold. What is already clear is narrower than the headlines: a machine can now produce, in hours, arguments that take people months to read.

What it means for you

Governments, funders, universities and ordinary readers will all meet this pattern again in other fields, and they will meet it first as a headline.

Four questions for the next “AI solved it” headline

1. Who checked it, and how? Lean, independent referees, or only the lab that made it. In this release, roughly half came without a Lean check, and three of 722 were withdrawn within a day.
2. What exactly was solved? Read the statement, not the headline. “Solved Navier–Stokes” does not mean water can move infinitely fast. It means that, under stated conditions, the equations stop being a valid model of a fluid.
3. What is withheld? Here the model, its training data and most of the process are not public, so nobody outside the company can repeat the run.
4. Who is the source? A lab announcing its own result, a university announcing a colleague’s, and a referee reporting an error carry different weight, and none of them is automatically wrong.

The world, and one reader

For institutions, the scarce resource is no longer the answer but the capacity to check it. Funders such as the Simons Foundation, journals and universities now decide what counts as progress, while the labs set the pace. We looked at the same problem of work nobody can inspect in When 1200 AI agents had a plan, and at what is missing in the public response in Is AI out of control already?.

For one reader, the practical skill is the same in mathematics, medicine or money: when results can be produced faster than anyone can read them, your edge is knowing which ones to trust and why. The four questions cost two minutes.

The rainforest is still there. What has changed is that the map is being drawn faster than anyone can walk it.

Factbox: AI and mathematics, as of 24 October 2026

Facts and figures change quickly. We do our best to get them right, so please check the sources below before relying on them; we hold no responsibility for decisions made on them. Published 24 October 2026.

Sources opened in full: OpenAI, On the Navier–Stokes Millennium Prize Problem (8 September 2026) · OpenAI, Sharing AI progress in mathematics (6 October 2026, updated 15 October) · OpenAI math repository README · Retraction Watch, OpenAI withdraws three preprints (8 October 2026) · Microsoft Research, the KLS conjecture (21 October 2026) · Wikipedia, OpenAI bulk release of mathematical papers

Also consulted, as excerpts only: TorontoToday, 17 October 2026 · Fortune, 7 October 2026 · EconoTimes on Terence Tao · Xie, catalogue of the release, arXiv, October 2026

© 2026 Think-Smarter / Leisure Media Group Ltd. All rights reserved.