Skip to main content
Back

Anthropic's Claude Helps Claim a Proof of the Riemann Hypothesis

Two frontier labs now claim proofs of Clay Millennium Prize problems, with Anthropic's Riemann hypothesis result arriving as Lean code awaiting independent checks.

  • Anthropic's Levent Alpöge posted a Lean-formalized proof of the Riemann hypothesis developed with Claude, awaiting verification.
  • The claimed method also resolves the Generalized Riemann Hypothesis, with the Extended version flagged as future work.
  • The artifact is millions of lines of Lean code corresponding to a proof Alpöge says runs over 250,000 pages.
  • OpenAI's September Navier-Stokes claim used roughly 10,000 coordinating agents on an unreleased internal model over 88 hours.
  • Clay Mathematics Institute requires published solutions to withstand two years of scrutiny before awarding any Millennium Prize.
  • Open Millennium problems remaining: P vs NP, Hodge, Yang-Mills mass gap, and Birch and Swinnerton-Dyer.

Levent Alpöge, a mathematician on Anthropic’s technical staff, has posted a claimed proof of the Riemann hypothesis developed with Claude. Roughly two months earlier, OpenAI announced a model-generated result related to the Navier-Stokes existence and smoothness problem. Both claims include Lean formalizations, and both remain unverified.

A short note, a vast formal proof

The Riemann hypothesis, open since 1859, concerns the zeros of the Riemann zeta function and their connection to the distribution of prime numbers. It states that every nontrivial zero has real part one-half. A confirmed proof would resolve one of the Clay Mathematics Institute’s seven Millennium Prize Problems.

On his personal site, Alpöge writes, “I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis.” His human-readable note runs four pages, while the reported formal proof spans millions of lines of Lean code. He estimates that a conventional rendering would occupy more than 250,000 pages.

Lean is a proof assistant whose kernel checks whether each proof term follows from declared definitions, theorems, and axioms. That process can verify an argument too large for line-by-line human review, provided the formal statement matches the intended theorem and the project stays within an accepted set of assumptions.

The claimed method also extends to the Generalized Riemann Hypothesis, according to Alpöge, and he expects further work to reach what he calls the Extended Riemann Hypothesis. GRH applies the same broad claim about zeros to a larger family of functions called L-functions. Because terminology varies across the literature, reviewers will need to compare the exact Lean declarations with each named conjecture.

Claude’s reported mathematical move

Alpöge credits Claude with proposing a key construction late in the project. For a sum containing X terms, the model suggested a chain of theta lifts with length proportional to √X. Theta lifting transfers number-theoretic information between related automorphic objects, allowing a difficult expression to be studied through a sequence of more structured ones.

The four-page note also contains a research timeline assembled by Claude, which Alpöge cautions may require correction. His account provides the current evidence for the division of work between mathematician and model. A successful Lean audit could establish the formal result, while the model’s precise contribution would still depend on research logs and Alpöge’s documentation.

OpenAI’s forced-fluid construction

On September 8, 2026, OpenAI published a claimed finite-time singularity construction for the Navier-Stokes equations, which describe fluid motion. The company says roughly 10,000 coordinating agents worked for about 88 hours using an unreleased internal model. Agents were upgraded when a further-trained version became available during the run.

OpenAI’s account says the construction shows a fluid solution becoming singular in finite time. Its scope includes versions of the equations driven by an external force. That distinction affects its relationship to the Millennium problem, whose best-known formulations focus on global smoothness or breakdown under specified conditions. OpenAI has said it will not seek the prize, and the official problem remains open.

Reported differences between the two AI-assisted projects
Dimension Riemann claim Navier-Stokes claim
Reported workflow One mathematician working iteratively with Claude About 10,000 coordinating agents
Reported model role Proposed a key theta-lift construction and helped document the research Searched for and formalized a forced-fluid singularity construction
Formal artifact Millions of lines of Lean Lean formalization
Current status Unverified Unverified; no prize claim planned

What a valid Lean audit requires

Independent reviewers must establish more than whether the repositories compile. Lean checks the theorem encoded in the project, so an audit must connect that formal declaration to the mathematical statement described in the announcement.

  1. Match the target theorem. Reviewers must verify every quantifier, domain, regularity condition, and definition in the top-level declaration.
  2. Inspect the axiom footprint. Commands such as #print axioms can expose dependencies on sorryAx, custom axioms, classical principles, or other assumptions requiring review.
  3. Reproduce the build. A credible release needs complete source code, pinned Lean and library versions, dependency hashes, build instructions, and realistic resource requirements.
  4. Map the trust boundary. Reviewers must identify generated certificates, external solvers, native computations, opaque binaries, and any step whose correctness is assumed outside Lean’s kernel.
  5. Review the mathematics. Experts still need to understand the argument’s structure, confirm that definitions reflect standard usage, and determine whether the claimed extensions follow.

Clay’s prize rules add a separate institutional process. A proposed solution must appear in a qualifying publication, at least two years must pass after publication, and the work must achieve general acceptance in the mathematics community before the institute considers an award.

The official scoreboard stays put

Status of the seven Millennium Prize Problems
Problem Official status Relevant result or claim
Poincaré conjecture Solved Grigori Perelman, 2003
Navier-Stokes existence and smoothness Open OpenAI forced-case claim, unverified
Riemann hypothesis Open Alpöge with Claude, unverified
P versus NP Open None
Hodge conjecture Open None
Yang-Mills existence and mass gap Open None
Birch and Swinnerton-Dyer conjecture Open None

The research workflow under test

When a model produces millions of lines of Lean, reviewers can delegate local logical checks to the kernel while concentrating on the theorem statement, assumptions, dependency graph, and reproducibility. Line counts, agent counts, and runtime describe computational scale; correctness depends on the formal target and its trusted foundation.

The two claims document a model-assisted workflow in which AI systems propose constructions, researchers direct the search, and a proof assistant checks the resulting derivations. Independent teams can evaluate that workflow once they can rebuild the artifacts, inspect their axioms, and match the top-level declarations to the intended mathematics.