Anthropic Mathematician Claims Claude Helped Prove Riemann Hypothesis

9 Min Read
Anthropic Mathematician Claims Claude Helped Prove Riemann Hypothesis

A famous claim, with checking still under way

Levent Alpöge, a mathematician at artificial intelligence company Anthropic, said on November 20 that he had discovered a proof of the Riemann hypothesis with help from the company's Claude model. The claim concerns a problem first posed in 1859 and tied to the pattern of prime numbers. If the proof holds up, it would resolve one of mathematics' best known open questions.

Contents
  1. A famous claim, with checking still under way
  2. What does the Riemann hypothesis say?
  3. What exactly is new in the November claim?
  4. How can Lean check a proof?
  5. Why is a readable explanation still needed?
  6. How did the announcement unfold?
  7. What would acceptance require?
  8. What to Know

By November 21, computer checks had provided substantial support for the formal proof's logic. Massachusetts Institute of Technology mathematician Andrew Sutherland said that 175,187 modules in the Lean version had been processed, and that checks by Lean's kernel and a separate program called nanoda had passed. Still, researchers were continuing tests of whether the code's definitions match the intended hypothesis, while mathematicians worked to understand the argument. The Clay Mathematics Institute continued to list the problem as unsolved.

The distinction matters: a computer can check whether a formal argument follows from its stated rules, while specialists must also establish that those rules express the famous problem and understand what the argument does. The claim is therefore an extraordinary development, not yet a prize winning resolution accepted by the mathematical community.

What does the Riemann hypothesis say?

Prime numbers are whole numbers greater than 1 that can be divided evenly only by 1 and themselves: 2, 3, 5, 7 and 11 are examples. They become less frequent among larger numbers, but their exact pattern is difficult to predict. Mathematicians can estimate how many primes lie below a chosen number, yet the estimate differs from the actual count.

Bernhard Riemann's 1859 work connected that pattern to the zeros of a mathematical function called the Riemann zeta function. A zero is an input at which a function's value is zero. The hypothesis says that every nontrivial zero of this function lies on one vertical line in the complex plane, where the real part is one half. A complex number has a real part and an imaginary part, so the line fixes one coordinate while the other can vary.

The idea is not that primes stop being irregular. Rather, the location of the zeros places tight limits on how much the number of primes can stray from a mathematical estimate. Many results in number theory rely on the hypothesis or would become stronger if it were proved.

The Clay Mathematics Institute named it one of seven Millennium Prize Problems in 2000 and set a US$1 million prize for an accepted solution. Its Millennium problem paper describes the connection between zeta zeros and prime counts. Earlier computations have checked very large collections of zeros, including more than 300 million in selected ranges, but checking examples cannot establish that every zero follows the rule.

Advertisement

What exactly is new in the November claim?

Alpöge's announcement goes beyond Anthropic's earlier result, made public in August. In that separate effort, an unreleased research version of Claude raised the guaranteed minimum share of zeta zeros known to lie on the critical line from 41.6% to 67.2%. That was a substantial improvement, but it did not show that all nontrivial zeros lie there, which is what the Riemann hypothesis requires.

Anthropic's August research post said Claude had not proved the hypothesis. It described a result about a related question: the fraction of zeros that can be shown to lie on the line. The move from 41.6% to 67.2% is an increase of 25.6 percentage points. Even 67.2%, however, leaves the question of the remaining zeros open.

That earlier work drew on published mathematics, and Anthropic mathematicians Levent Alpöge and Ralph Furman examined it. The company also said specialists Brian Conrey and Dan Goldston reviewed the work. Claude produced a formal Lean proof of that lower bound, which passed Lean's validation tool. Anthropic reported that the effort used 31 million output tokens across two sessions and about 60 subagents, or AI processes working on separate tasks.

The current claim is different in scope. Alpöge says that he and Claude proved the full hypothesis, and that their method also proves a broader version, the Generalized Riemann Hypothesis, which concerns related functions. That broader claim, too, requires careful checking. A proof of either statement could affect many mathematical results that now depend on those hypotheses being true.

How can Lean check a proof?

Lean is a programming language and proof system in which a theorem and the steps used to prove it are written in a precise form. Its small checking core tests whether each step follows from earlier steps and the stated foundations. That makes it harder for a gap in a long chain of reasoning to pass unnoticed than it might be in ordinary prose.

According to the reported verification, the Lean project included five theorems, among them the Riemann hypothesis as defined in the Mathlib mathematical library and a version concerning Dirichlet L functions. Sutherland said the modules passed checks using Lean's kernel, the program that checks proofs, and the independent program nanoda. The work was reported to use only Lean's three standard axioms, basic rules accepted by the system, rather than adding new ones.

Sutherland described the independent kernel check in a statement about the Lean proof. He said the remaining risk of a hidden flaw looked extremely small:

The Lean proof has been compiled and independently rechecked by the Lean kernel with only the three standard axioms. Honestly the chance that anything is fishy here is approximately zero.

That is strong evidence about the formal code, but a second issue remains: does the theorem represented by the code match the Riemann hypothesis as mathematicians understand it? Imperial College London mathematician Kevin Buzzard urged scrutiny of the statement itself, as well as the proof. A comparator run was still pending at the reported November 21 checkpoint. Such a test can look for altered definitions or other changes that might make a computer verify a different claim.

Advertisement

Why is a readable explanation still needed?

The public account is not yet a conventional mathematical paper that specialists can read from beginning to end. Alpöge released a short note of about four pages outlining the claim. He has said that a full written proof would run to more than 250,000 pages, while reporting on the announcement described the Lean formalization as extremely large. A compact outline and a computer checked formalization serve different purposes: the first can describe the idea, while the second records detailed logical steps.

Researchers still need to understand the mathematical route through those steps. A formal proof can establish that a conclusion follows from its definitions and rules without making its central insight easy for people to see. The work of specialists is to check the statement, examine the argument's structure and explain which ideas led to the result.

Oxford number theorist James Maynard said the result would be a historic milestone if it fully checks out. Andrew Granville, a number theorist at the University of Montreal, criticized the released note as difficult to follow and called for a clearer account. Terence Tao, a leading mathematician, said the process of understanding the proof would take time. Their differing reactions reflect two separate tasks: establishing that the formal code proves the claimed theorem and making sense of the mathematics behind it.

Machine checking can shorten the search for a logical error, but it does not automatically settle what a theorem means or why it matters. Researchers will need to connect the formal statement with the accepted formulation of the hypothesis and produce an explanation that the field can evaluate.

How did the announcement unfold?

The recent history shows how the new announcement builds on, but is distinct from, Anthropic's earlier work:

  • August 10, 2026: Anthropic announced Claude's improvement to the known minimum share of zeta zeros on the critical line, from 41.6% to 67.2%. The company updated its account on August 13.

  • November 19: Anthropic made its stock market debut.

  • November 20: Alpöge announced the claimed proof, saying Claude had become central to the work near its end.

  • November 21: Researchers reported extensive Lean checks, while a comparator run and mathematical review remained under way. The Clay Institute still listed the hypothesis as unsolved.

The timing drew attention because the announcement came one day after Anthropic's listing. Alpöge said he had discussed the timing with chief executive Dario Amodei and acknowledged that observers might view it cynically, while rejecting that interpretation. The timing does not answer the mathematical questions, which turn on the statement and proof.

Advertisement

What would acceptance require?

The Clay Institute's prize rules require a proposed solution to be published in an eligible peer reviewed outlet and to gain broad acceptance among mathematicians. The solution must then remain under scrutiny for at least two years after publication before the institute considers awarding the prize. That process gives researchers time to check the proof, identify mistakes and assess whether it establishes the stated problem.

At the November 21 checkpoint, the claim had not completed those steps. No decision date for the prize was given. The immediate work is technical and mathematical: finish the independent code checks, confirm that the formal theorem is the Riemann hypothesis, and prepare an account that researchers can inspect and discuss.

The claim arrives amid a wider burst of AI generated mathematical work. Other companies have announced solutions or progress on difficult problems, some of which still need review. The volume creates a practical challenge: researchers must decide which claims merit attention and spend time checking them. In this case, Lean offers a way to test a huge number of logical steps, while human mathematicians remain central to judging the claim's meaning and significance.

What to Know

  • Anthropic mathematician Levent Alpöge said on November 20 that Claude helped him prove the Riemann hypothesis.

  • The hypothesis says every nontrivial zero of the Riemann zeta function lies on a line with real part one half.

  • Lean checks reportedly covered 175,187 modules, but a comparator run and wider mathematical review were still in progress on November 21.

  • Anthropic's August result raised the known minimum share of zeros on the line from 41.6% to 67.2%; it was not a proof of the full hypothesis.

  • The Clay Mathematics Institute continued to list the problem as unsolved. Its prize rules require publication, broad acceptance and at least two years of scrutiny.

Share This Article

You May also Like