Top 5 in AI

Signals

Did AI Just Prove the Riemann Hypothesis? What Claude and an Anthropic Mathematician Claim, What Has Been Checked, and Why AI Labs Chase Math Problems

By the Top5Apps editorial team · Published November 21, 2026 · Updated November 21, 2026 · 13 min read

Share
A looping orange and blue curve drawn on a pair of axes over pale concentric rings. The curve starts on the horizontal axis to the left of the center and swings through the center point again and again in widening loops.
The curve traces the values of Riemann's zeta function along the line named in the hypothesis. Each pass through the center is a zero. Still from Quanta Magazine's video The Riemann Hypothesis, Explained. Credit: Emily Buder/Quanta Magazine; Guan-Huei Wu and Clay Shonkwiler for Quanta Magazine. Shown editorially.

On November 20, 2026, Levent Alpöge, a mathematician who works at Anthropic, posted a four-page note that says he and Claude have proved the Riemann Hypothesis. Mathematicians have tried to prove it since 1859, and many of them call it the most important open problem in their field. Alpöge's note opens with the claim: 'I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis.'

(Disclosure, per The Receipts Standard: this site is produced with Claude-family tools, and the claim covered here comes from an Anthropic employee working with Claude. Every fact below comes from a page we read on November 21, 2026, and we link each one.)

The short answer is that computers have checked the proof and people haven't read it yet. Alpöge released the argument as code in Lean, a language built so that software can check every step of a proof. By the morning of November 21, an MIT mathematician and the Lean community had each run that code through independent checking programs, and those programs accepted it. No mathematician has yet read and understood the argument from start to finish, and Alpöge says he hasn't fully mastered it himself.

This post explains what the problem is, what has been checked, whether the proof came from brute force or from ideas, and why AI companies keep aiming their models at math.

What the Riemann Hypothesis says, in plain English

Prime numbers are the whole numbers that only divide by 1 and themselves, such as 2, 3, 5, 7 and 11. They get rarer as numbers get bigger, and they don't follow an obvious pattern. Mathematicians have a formula that estimates how many primes sit below any number you pick.

The Riemann Hypothesis says that this estimate is about as accurate as it could possibly be. Bernhard Riemann proposed it in 1859. In technical terms, he connected the primes to a function called the zeta function and guessed that all of its important zeros sit on one straight line.

Quanta Magazine's video explainer, presented by the Rutgers mathematician Alex Kontorovich, walks through the idea in about 16 minutes.

Quanta Magazine's explainer from 2021, presented by Alex Kontorovich of Rutgers University.

Scientific American calls it 'the most famous open problem in math' and notes that the Clay Mathematics Institute has offered $1 million for a proof. The same article says progress was so rare that most top experts spent little time working on it.

Alpöge's code goes further than Riemann's original statement. It also covers a larger family of related functions called Dirichlet L-functions, which is one version of what mathematicians call the Generalized Riemann Hypothesis.

What has been checked so far

Alpöge didn't publish a paper that a person can read. His note says he chose 'to only present Lean,' and it links to 14 archive files of code. A program called the Lean kernel reads code like this and accepts it only if every step follows from the step before.

The Lean community copied the code into a public repository and measured it. Their notes count 175,187 files and 17,374,215 lines, with 818,755 separate theorems and lemmas. Alpöge's note says a printed version runs past 250,000 pages.

File, line and theorem counts are from the Lean community's verification notes. The page counts are from Alpöge's note and Nature.

Andrew Sutherland, a mathematician at MIT, published his own verification files. His page says the code compiles with no errors, and that two independently written checking programs, called nanoda and con-leche, both accept the full proof. A separate set of receipts from an independent GitHub user reports the same result.

Sutherland's page also answers the question a careful skeptic asks first: did the code prove the real Riemann Hypothesis, or a look-alike statement? The page says the proof targets the definition of the Riemann Hypothesis that already existed in Mathlib, the shared math library that Lean users maintain. Alpöge and Claude didn't write their own version of the statement.

Sutherland added a second test that avoids that definition altogether. The Riemann Hypothesis is known to be equivalent to a plain statement about divisors called Robin's inequality. His page combines Alpöge's code with someone else's Lean proof of that equivalence, and the checkers accept the divisor statement too.

QuestionStatus on November 21Source
Did Alpöge publish the claim and the code?YesHis note and 14 archive files
Does the code compile without errors?Yes, all 175,187 filesMIT page, Lean community repository
Does it use any shortcuts or unproven assumptions?No. It relies only on Lean's three standard axiomsMIT page
Do independent checkers accept it?Yes: nanoda, con-leche and lean4leanMIT page
Is the statement the standard one?Yes. It is Mathlib's own definitionMIT page, Lean community notes
Are all checks finished?No. Two slower runs were still in progressMIT page, updated 12:55 UTC
Has a person read and understood the argument?NoAlpöge's note
Is there a readable paper?No. Alpöge says he intends to write oneAlpöge's note
Status as of the morning of November 21, 2026. Sutherland's page says his checks were run with Claude Code, Anthropic's coding tool. The checking programs themselves were written by other people.
Our summary of Andrew Sutherland's verification page at MIT and the Lean community's repository, read November 21, 2026.

Sutherland states the limit of all this in one sentence. A passing check 'relies on those kernels and on Mathlib's definitions being what they are intended to be.' Several independent kernels now agree, which makes a software bug an unlikely explanation.

Mathematicians quoted in the press are careful about the next step. James Maynard, a Fields Medal winner at Oxford, told Nature: 'If this fully checks out, it is clearly a historic milestone in mathematics.' Terence Tao wrote on his blog: 'I imagine the digestion process will take some time, though in this case at least we have a human author to communicate with to assist with this task.'

Sutherland went further after he compiled the code himself. He told Scientific American: 'This is indeed an historic moment.' He added that 'there is much work to be done to understand the mathematical details of the proof.'

Is it brute force, or real creativity?

The early evidence points to something between the two. Experts who have looked at the code describe a huge combination of existing mathematics, steered by a person, with key steps that the person credits to the model. Nobody outside Anthropic can confirm that account yet.

Brute force in the strict sense can't produce this proof. A checker such as Lean only accepts or rejects what it's given, and nobody finds 17 million valid lines by trying random ones. The size of the proof does show that a model can now sustain a very long argument without losing the thread. That part is our reading of the evidence.

Alpöge's note describes a division of work. He writes that he had planned for this problem since his student days, that he combined two ideas in a simpler test case, and that he 'guided us to the right situation.' After that, in his words, 'Claude far outpaced me near the finish line.' He says 'tens' of key steps came from Claude, and he gives one example that he expects will make mathematicians 'fall out of their chairs.'

Both columns paraphrase Alpöge's note of November 20, 2026. Nobody outside Anthropic has confirmed who contributed what.

He is also frank about what he doesn't understand. He writes that he is 'presenting the world with a complete argument before having mastered' the mathematics around it, and that for some parts he is 'paraphrasing' a summary from Claude that 'might make zero sense.' He expects the argument 'will be massively simplified.'

Three early reactions from mathematicians help answer the creativity question.

  • Terence Tao wrote on his blog that the strategy, 'from very initial analysis, appears to use the full power of modern developments' in two large areas of number theory. In a comment he added: 'The RH proof uses a century of past literature on related topics.'
  • Timothy Gowers posted: 'Who would have thought that the convex hull of known mathematics contains such riches?' We read that as saying the proof seems to be built by combining things mathematicians already knew.
  • Daniel Litt posted: 'Seems like I was wrong about this!' He called his mistake 'largely a reflection of what theory I thought would need to be built to prove RH.'

Those reactions describe a kind of creativity that people rarely manage: holding a hundred years of results from several fields in view at once and finding the combination that works. Jordan Ellenberg added that Alpöge appears to have 'a mental picture of the whole thing' in outline.

Other mathematicians are less comfortable. Harald Helfgott wrote in the Lean community's discussion thread that Alpöge 'has both great resources at his disposal and a rather limited understanding of what the tools have done, by his own admission.' Andrew Granville told Nature: 'What he released is garbled.'

Two facts are still missing. Anthropic hasn't said which model did the work, and the BBC reports that it was 'a non-public version' of Claude. Nobody has said how much computing it took. Tao made the same point in a comment on his blog: 'We don't know how much compute in aggregate has been spent.'

Why AI companies keep aiming their models at math

The Riemann claim is the largest in a run of math results from AI labs, and the pace picked up sharply in 2026.

Dates are from each company's own announcement or the author's post. Several of the 2026 results are still under review by mathematicians.

Anthropic had aimed Claude at this exact problem before. In August it reported that a staff member gave an unreleased research version of Claude 'an unreasonable challenge' to attempt the Riemann Hypothesis. Claude didn't succeed, and it improved a related result along the way. Anthropic wrote then: 'We don't expect that the techniques Claude used will lead to proving the Riemann hypothesis.'

The companies give four reasons in their own words, and critics add a fifth.

1. Math can be checked

Most AI output is hard to grade. A proof is either valid or it isn't, and software can decide. OpenAI wrote in May: 'Mathematics provides a particularly clear testbed for reasoning: the problems are precise, potential proofs can be checked, and a long argument only works if the reasoning holds together from beginning to end.'

2. The old tests ran out

Labs measure models with benchmark tests, and OpenAI says its models outgrew the math tests it had. Its notes for the 700-plus math manuscripts it released in October say: 'We expanded these evaluations after performance on our existing mathematical evaluations saturated.' An open problem can't be memorized, because no answer exists to memorize.

3. Math is practice for science

The labs say the real target is research in general. OpenAI's May post calls its math work 'part of our longer-term path toward more automated research' in fields such as biology, physics and medicine. Demis Hassabis of Google DeepMind describes choosing 'root node problems that, if you solve them, you unlock entire new branches of research.'

4. Checked proofs make the models better

Writing proofs in Lean gives a model instant feedback on whether each step holds. Anthropic said in September, when Claude produced a computer-checked proof of Fermat's Last Theorem: 'Writing Lean also seems to help Claude prove novel results.'

5. It is very good publicity

A famous solved problem tells customers and investors that a model can reason, in a way no benchmark chart can. The labs watch each other closely. When OpenAI published a Navier-Stokes result in September, Sam Altman wrote that 'we tried this because there were rumors on the internet last week that Anthropic's models had solved a millennium problem and we were curious if ours could do it too.'

The Riemann claim arrived at a sensitive moment for Anthropic. The company priced its initial public offering at $83 a share and began trading on November 19. NPR reported a valuation of about $1.5 trillion and a first-day close of $97.17. Alpöge posted his note the next day, and the shares fell 9% that day in a wider sell-off of AI stocks.

Alpöge's note addresses this directly: 'The precise timing is horrible.' He writes that he messaged Anthropic's chief executive late on Sunday, that 'there was a constraint to respect,' and that the 'obvious cynical view, to do with the timing' is false. Nature reports that Granville cites the fall in the company's stock value as the reason for the timing. We found no statement about the proof on Anthropic's news page or its X accounts when we checked on November 21.

Left column from OpenAI, Anthropic and Google DeepMind pages. Right column from the September 2026 declaration at mathandai.org and from Nature's reporting.

Many mathematicians object to the whole race. In September, 28 Fields Medal winners signed a declaration that says 'the push by AI companies to solve mathematical problems as a benchmark is detrimental to the science of mathematics.' The signers include Tao and Maynard. The declaration argues that 'solving problems is only a tool and proxy for achieving the primary goal of conceptual understanding and insight.'

Why it matters whether the proof holds

For mathematics, a correct proof would turn a large body of conditional results into established ones. Over the decades, mathematicians have proved many theorems that begin 'if the Riemann Hypothesis is true.' Scientific American describes the generalized version as 'instantly unlocking a deluge of truths' that depend on it. Hector Pasten, a mathematician quoted there, compares a correct and understood proof to 'something like a Rosetta Stone' between branches of math.

For daily life, almost nothing changes right away. The same article is clear about this: the result 'won't upend global banking or make the internet run faster.' A proof would confirm a pattern in the primes that mathematicians have long assumed is real.

For AI, the result would be the hardest problem a model has helped solve. Scientific American's reporter describes it as one that the smartest people on the planet 'all agreed they were nowhere close to figuring out.'

The $1 million prize is a separate question with a slow clock. The Clay Mathematics Institute's rules require a solution to be published in a qualifying outlet, to wait at least two years, and to receive 'general acceptance in the global mathematics community.' A Lean file with no readable paper doesn't meet the first condition yet.

How this could help the rest of us

The useful pattern here is a model that writes long, difficult work and a separate program that checks every step. That pattern already runs outside pure math.

  • Agents have checked a messaging protocol. Google DeepMind said on November 10 that its AlphaProof Code agents wrote more than six million lines of Lean to prove security properties of the code behind the Signal messaging protocol, in four weeks.
  • Microsoft checks encryption code this way. Microsoft described in July how it uses Lean proofs for the cryptography code used across Windows and Azure. The post says: 'An agent may propose a proof script, but Lean independently verifies whether the proof is valid.'
  • Claude has verified a common software library. Lean's creator, Leonardo de Moura, reported in February that Claude converted a widely used compression library to Lean and proved that compressing and then decompressing a file returns the original.

If models can prove 800,000 lemmas about prime numbers, they can also prove that a banking app never loses a transaction or that a medical device never skips a safety check. Google DeepMind says work like its Signal proof 'historically takes expert teams years.' The banking and medical examples are our inference, and none of those companies has promised them.

There are limits to how far the pattern stretches. Math has a checker and most other work doesn't. No program can confirm that a business plan, a diagnosis or a news summary is correct. In those areas a person still has to do the checking, and the Riemann proof shows how large that job can get: Sutherland's machines compiled 17 million lines in just over an hour, and people will need months or years to understand them.

The model that did this work also isn't available to the public. If you use Claude, ChatGPT or Gemini today, you aren't using the system that produced this proof.

The practical lesson for anyone who works with AI is the same one this proof depends on. Ask the model for work that something else can check, such as a test, a calculation or a source you can open. We wrote a guide to doing that with code in How to get Claude to write better code.

What we still don't know

  • We don't know whether mathematicians will accept the argument. Machine checks have passed. Acceptance by mathematicians depends on a readable proof, and Alpöge says he intends to write one.
  • We don't know which model did it, or at what cost. Anthropic hasn't named the model or the amount of computing involved.
  • We don't know how much of the reasoning was Claude's. The only account comes from Alpöge, who works for Anthropic.
  • The last checks haven't finished. Sutherland's page listed two slower runs as still in progress at 12:55 UTC on November 21.
  • The proof may get much shorter. Alpöge expects a much simpler version, and members of the Lean community are already using AI tools to translate the code into readable mathematics.

The short answers

  • Did AI prove the Riemann Hypothesis? Alpöge claims that he and Claude did. Independent software checks accept the proof. Mathematicians haven't read and accepted it yet.
  • Why does it matter? It would settle the most famous open problem in math, confirm many results that depend on it, and show that an AI model can finish a problem experts thought was out of reach.
  • Was it brute force or creativity? Early expert readings describe a vast combination of existing mathematics, guided by a mathematician. That takes more than brute force, and nobody can measure the model's share yet.
  • Why do AI companies do this? Math is checkable, older tests have run out, the labs see math as practice for science, and a famous result is powerful publicity.
  • How does it help in the future? The same approach can prove that software is correct, and the labs say they want to apply it to science. The model itself isn't public.

We'll update this post when the remaining checks finish, when Anthropic or the Clay Mathematics Institute makes a statement, and when mathematicians publish their first readings of the argument.

Where these apps rank