Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseAPI / Product · ActionableHabr AI→ original

Mathematician Says Claude Helped Prove the Riemann Hypothesis: Code Verification Reaches 99%

Anthropic mathematician Levent Alpöge claims he proved the Riemann hypothesis with the help of Claude. The problem, posed in 1859, concerns prime numbers. The proof is written in Lean, which works like a compiler for mathematics: it accepts only reasoning without gaps. A public verification run covering 175 thousand files has reached 99% without errors, but there is no verdict yet. The lesson for developers is simple: a model's answer is more trustworthy when a strict machine checks it.

AI-processed from Habr AI; edited by Hamidun News
Mathematician Says Claude Helped Prove the Riemann Hypothesis: Code Verification Reaches 99%
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Key metric #1An enthusiast launched the verification on a server with 128 cores and 244 GB of memory
Key metric #2The build includes 14 archives, 228 MB of code and more than 26 thousand build tasks
Key metric #3Alpöge believes the approach could also work for the generalized Riemann hypothesis

Anthropic mathematician Levent Alpöge claims that he and Claude found a proof of the Riemann hypothesis, one of the best-known unsolved problems in mathematics. For now, this is a claim, not an accepted fact. But the story is worth reading for its verification method, not for the sensation. The proof is not a persuasive text. It is code in the Lean language, which a computer checks on its own. This is the same principle behind reliable AI work in software development: the model proposes, and a strict external system accepts or rejects.

What exactly was claimed and why it matters

The German mathematician Bernhard Riemann formulated the hypothesis in 1859. In the 167 years since, the best mathematicians in the world have not managed to prove it. It is one of the seven Millennium Prize Problems, and the Clay Mathematics Institute promises a million dollars for solving each of them.

In the simplest terms, the hypothesis concerns the distribution of prime numbers, those divisible only by one and themselves. Among the natural numbers they look scattered at random, but the Riemann hypothesis describes a strict pattern behind that chaos. Much depends on it: a huge number of results in number theory are proven with the caveat "if the Riemann hypothesis is true." A real proof would remove that caveat from a multitude of theorems at once.

In the quarter century since the prizes were established, only one of the seven problems has been solved: Grigori Perelman proved the Poincaré conjecture in 2002–2003. In 2010 he was awarded a million dollars, but he refused the money, just as he had earlier refused the Fields Medal.

How Lean differs from "a beautiful text"

Imagine two ways to check someone else's code. In the first, you read it and decide that everything looks reasonable. In the second, you run the compiler and the tests: they either pass or they don't. Lean is the second way, applied to mathematics. Every step of the reasoning is written as code, and the computer checks it sequentially. A model can confidently write any nonsense, but in Lean an unverified step simply won't build.

The scale of the verification is impressive. One enthusiast launched a public build on a server with a 128-core processor and 244 GB of RAM. The project has 14 archives, 228 MB of source code, more than 175 thousand Lean files and over 26 thousand build tasks. At the published stage, verification reached 99% with no errors detected.

Why it is too early to declare victory

A build that is 99% error-free does not mean the hypothesis is proven. Two main questions remain.

First: does the final theorem in Lean actually match the original problem? A software analogy: the tests can all be green while checking something other than what the specification required. If the Lean statement differs from the real Riemann hypothesis, the computer will faithfully confirm the wrong theorem.

Second: does the proof rely on inadmissible assumptions, that is, statements accepted without justification? Independent specialists will have to check this.

What to take away in practice

First, if the result holds up, we will get a rare chance to assess AI's contribution to a real scientific discovery, not to a result on yet another benchmark. A human and a language model working together on a problem that stayed open for more than a century and a half is on a completely different scale from model leaderboards.

Second, the pattern itself is useful for engineers: the stricter the external checker (a compiler, tests, a formal specification), the safer it is to hand long tasks to a model. What deserves trust is the verification, not the confident tone of an answer.

Third, Alpöge says the approach may also work for the generalized Riemann hypothesis, an even broader mathematical problem. That, too, is so far only a claim.

It is too early to say the Riemann hypothesis has been solved. But if the proof withstands independent verification, 2026 could go down in history not only for mathematics but for artificial intelligence as well. You can follow the public verification at the link from the original source: http://106.14.11.164.

Story

Key Questions Answered

What exactly was claimed and why it matters?

The German mathematician Bernhard Riemann formulated the hypothesis in 1859. In the 167 years since, the best mathematicians in the world have not managed to prove it. It is one of the seven Millennium Prize Problems, and the Clay Mathematics Institute promises a million dollars for solving each of them. In the simplest terms, the hypothesis concerns the distribution of prime numbers, those divisible only by one and themselves.

How Lean differs from "a beautiful text"?

Imagine two ways to check someone else's code. In the first, you read it and decide that everything looks reasonable. In the second, you run the compiler and the tests: they either pass or they don't. Lean is the second way, applied to mathematics. Every step of the reasoning is written as code, and the computer checks it sequentially. A model can confidently write any nonsense, but in Lean an unverified step simply won't build. The scale of the verification is impressive.

Why it is too early to declare victory?

A build that is 99% error-free does not mean the hypothesis is proven. Two main questions remain. First: does the final theorem in Lean actually match the original problem? A software analogy: the tests can all be green while checking something other than what the specification required. If the Lean statement differs from the real Riemann hypothesis, the computer will faithfully confirm the wrong theorem. Second: does the proof rely on inadmissible assumptions, that is, statements accepted without justification?

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…