Anthropic mathematician says Claude proved the Riemann hypothesis, but published only unreadable Lean code
Imagine a code review of 250,000 pages. You cannot read it, so the only thing you can trust is a green test suite. That is roughly the situation Anthropic mathematician Levent Alpöge created when he claimed that, working with Claude, he proved the Riemann hypothesis about prime numbers, but published only Lean code: a program that checks every logical step. Andrew Sutherland of MIT compiled the code, but the final verification is not finished, so for now this is a claim, not an accepted fact.
AI-processed from TNW; edited by Hamidun News
Imagine someone sends you a pull request of 250,000 pages. You cannot read it, so you look at one thing: do the tests pass? Mathematics now finds itself in roughly this situation. Levent Alpöge, a mathematician at Anthropic, wrote on his website that he and the Claude model proved the Riemann hypothesis. It is the most famous unsolved problem in mathematics and one of the seven Clay Mathematics Institute "Millennium Prize Problems," with a $1 million prize. There is no readable paper for the proof. There is only code, which a machine checks.
What is claimed, and where Lean fits in
Bernhard Riemann formulated the hypothesis in 1859. It describes how prime numbers (2, 3, 5, 7, 11, and so on) are distributed along the number line. Alpöge writes that his method also proves the generalized Riemann hypothesis, a much broader version.
The result was published as Lean code. Lean is a formal verification system that acts as a strict compiler: every logical step is written out explicitly, and if the code compiles, the chain of reasoning should contain no errors. According to Alpöge, a human-readable version would be enormous: his own draft exceeded 250,000 pages. He said publishing something like that would be an "abuse of the word slop," so he released only the Lean code.
The author acknowledges that he has not yet mastered all the mathematics surrounding the argument. He says Claude made the final steps. His note includes a timeline of the work from October 6 to November 16, which he claims Claude itself compiled. Alpöge promises to work through the argument, simplify it, and write a proper text.
What independent mathematicians say
Andrew Sutherland of MIT compiled the Lean proof independently on Friday. He told Scientific American that the chance of a trick is "approximately zero."
One automated check has not yet finished. It is a comparator, and the test analogy applies here too. It verifies that the code proves the Riemann hypothesis as originally stated, rather than a slightly altered statement. This is what happens when tests pass but check the wrong function. Until that comparison is complete, James Maynard of Oxford puts his confidence at "98 or 99 percent."
Terence Tao of UCLA warned that understanding the proof will take time. In his initial assessment, the argument draws on the full power of modern number theory.
Not everyone welcomes this way of publishing. Brian Conrad of Stanford called releasing a proof that its author does not fully understand "outright outrageous." His remarks were reported by the NYT.
Timing, competing work, and takeaways for practitioners
The claim came the day after an Anthropic announcement, at a time when the sell-off in AI-related assets was still gathering pace. Alpöge himself writes that the timing has an "obvious cynical reading," but he says that reading is false. The breakthrough came late on Sunday evening. He then wrote to Anthropic CEO Dario Amodei and asked about the announcement. There was a "constraint that needs to be respected."
That same day, American Mathematical Society president Ravi Vakil posted a different proof, written by mathematician Alessandro Masullo. According to Vakil, it is entirely human-made, without AI, and Conrad checked it over several months. About eight hours later, Vakil added that several people had found problems in it using AI models.
This is not Alpöge's first high-profile claim. In July he published a counterexample to the Jacobi conjecture, found by Claude. And on Friday, together with Tristan Bakmaster of NYU, he made an announcement about the Navier–Stokes equations.
What matters here for people building AI products:
- Verifiability matters more than readability. When a result cannot be read by eye, an automated check becomes the arbiter. For agents, this is a familiar logic: tests and formal verification are more reliable than a neat explanation.
- The checks themselves need checking. The comparator exists precisely to confirm that the problem actually proved is the one that was posed.
- Understanding lags behind the result. The author cannot explain his own proof, and the community will spend months working through it. The same will apply to AI-written code that works but that nobody understands.
Until the comparator finishes and mathematicians have worked through the ideas, this remains a claim, not an established fact. Whether to celebrate should wait for those checks.
- Nov 22, 2026Nasdaq 100 Falls 10.4% in a Week, Its Worst Stretch Since the Pandemic
- Nov 22, 2026Android work profile closes real-IP leak that Happ and v2rayNG don't catch
- Nov 22, 2026Anthropic Goes Public on Nasdaq: Shares Rise from $83 to $111, but Can Safety Hold Up?
- Nov 22, 2026Databricks Is in Talks for a Round at a $225 Billion Valuation, Nearly 20% Above August
Key Questions Answered
What is claimed, and where Lean fits in?
Bernhard Riemann formulated the hypothesis in 1859. It describes how prime numbers (2, 3, 5, 7, 11, and so on) are distributed along the number line. Alpöge writes that his method also proves the generalized Riemann hypothesis, a much broader version. The result was published as Lean code. Lean is a formal verification system that acts as a strict compiler: every logical step is written out explicitly, and if the code compiles, the chain of reasoning should contain no errors.
What independent mathematicians say?
Andrew Sutherland of MIT compiled the Lean proof independently on Friday. He told Scientific American that the chance of a trick is "approximately zero." One automated check has not yet finished. It is a comparator, and the test analogy applies here too. It verifies that the code proves the Riemann hypothesis as originally stated, rather than a slightly altered statement. This is what happens when tests pass but check the wrong function.
What are the key details on Timing, competing work, and takeaways for practitioners?
The claim came the day after an Anthropic announcement, at a time when the sell-off in AI-related assets was still gathering pace. Alpöge himself writes that the timing has an "obvious cynical reading," but he says that reading is false. The breakthrough came late on Sunday evening. He then wrote to Anthropic CEO Dario Amodei and asked about the announcement.
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.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.