Kivorent News

Sunday, November 22, 2026

Tech and AINews

Lean and Mathlib groups back the Claude-assisted Riemann Hypothesis proof

The Lean FRO and the Mathlib Initiative endorsed a joint statement on November 21 saying Levent Alpöge's machine-checked proof of the Riemann Hypothesis, built with Anthropic's Claude, holds up. The formal check is done; the human understanding is not.

Cortex, covering AI models, labs and agents
, 3 min read

175,187Lean modules reported to compile with zero errors

What to know

  • A joint statement endorsed by the Lean FRO and the Mathlib Initiative on November 21 certified the Lean proof, according to Unite.AI, Startup Fortune, AgentsPulse and Data Studios.
  • Reported checks: the final theorem is Mathlib's existing RiemannHypothesis definition, all 175,187 modules compile, and the main theorems use only Lean's three standard axioms.
  • No human-readable proof exists yet, Anthropic has made no formal statement, and the Clay Mathematics Institute still lists the problem as unsolved.

Days after a claimed proof of the Riemann Hypothesis appeared on the website of Anthropic mathematician Levent Alpöge, the groups that run the Lean proof assistant and its shared math library have backed it. A joint statement endorsed by the Lean FRO and the Mathlib Initiative on November 21 says the machine-checked proof, built with Anthropic's Claude, is correct, according to reports from Unite.AI, Startup Fortune, AgentsPulse and Data Studios. We could not open the statement itself on the Lean community blog, so the details below come from those reports.

When we first covered the claim, nobody outside Alpöge's circle had confirmed anything. That has changed for the formal part. It has not changed for the mathematics a human can read.

What was checked

The reports describe four checks, and they close the two gaps we flagged on Friday.

  • The statement. The final theorem is Mathlib's existing RiemannHypothesis definition, written and vetted before this claim, not a new statement from the proof's authors. Tools that compare a clean file importing only Mathlib against the proof's theorem found a match, which rules out symbols quietly redefined along the way.
  • The build. All 175,187 Lean modules compiled with zero errors on Lean and Mathlib 4.33.1, per a log kept by MIT mathematician Andrew Sutherland, who ran his checks with Claude Code, according to Unite.AI and AgentsPulse. Dozens of people compiled the library and reproduced the same export, Unite.AI and Startup Fortune report.
  • The axioms. The five headline theorems rest only on Lean's three standard axioms, reports AgentsPulse, Unite.AI and Kingy AI.
  • The kernels. The exported proof, about 61.6 GB, was run through Lean's own kernel and independent checkers written by other people, including nanoda and con-leche, which accepted it.

Some runs were still going at Sutherland's last reported update, including a full end-to-end comparison. The reports differ on which had finished and on whether the export has been matched against a fully sandboxed rebuild. Sutherland told MLQ that the chance anything is wrong is approximately zero, Startup Fortune reports.

What the certificate does not cover

A Lean check proves that the files prove the statement. It does not explain why. No human-readable write-up of the argument exists, and Alpöge says he is still working to understand and explain it, per Unite.AI and Kingy AI. Oxford's James Maynard told IBM he expects the proof to hold but that understanding it will take a lot of work.

The scope is also narrower than the first announcement. AgentsPulse and Kingy AI report that the generalized result covers Dirichlet L-functions only, with broader families left as future work.

Anthropic has issued no formal statement, and the note does not name the Claude model used. The Clay Mathematics Institute still lists the hypothesis as unsolved; its $1 million prize requires publication in a qualifying outlet, two years of scrutiny and general acceptance. Clay's president, Martin Bridson, called the announcement "of historic significance," AgentsPulse reports, citing the Associated Press.

Why it matters

A famous open problem has been accepted by the people who maintain the checker before any human could follow the argument. That moves the hardest question about AI in mathematics from whether a model's output is right to whether people can learn anything from it. Earlier work, such as Google's agents formalizing the Weil conjectures and DeepMind's AlphaProof Code, checked results people already understood. It also lands as Anthropic, newly public, is pitching its models on long, verifiable work, a claim that rarely gets a test this clean. For how to weigh lab claims generally, see our benchmark explainer.

Sources

AnthropicAI modelsAI agents

Written by an AI correspondent under standards set by editor Jacob Knight. Spotted an error? Tell us and we will fix it.