Anthropic mathematician says Claude helped prove the Riemann hypothesis in six weeks
Anthropic mathematician Levent Alpöge says he proved the Riemann hypothesis with Claude in six weeks. It is a prime-number problem nearly 170 years old. The proof is written in Lean code, so a computer checks each step, much as automated tests check a program. For now, it is one author's claim, and mathematicians are still working out whether the code describes the Riemann hypothesis itself. The takeaway for practitioners: trust AI where a machine verifies the result, not your eyes. The note appeared the day after Anthropic's IPO on Nasdaq. Italian mathematician Alessandro Massullo spent 12 years producing an incomplete proof without AI. The Clay Institute offers $1 million for solving each of the seven "Millennium Prize Problems."
AI-processed from Guardian; edited by Hamidun News
Imagine asking an AI to write a complex database migration. It looks convincing, but running it in production is scary: the model may have been confidently wrong somewhere. Now raise the stakes to one of the most famous unsolved problems in mathematics. That is where the story of the Riemann hypothesis now stands, as people try to confirm or refute it. Levent Alpöge, an American mathematician of Turkish origin who works with Anthropic, says that together with Claude he brought its proof to completion. This is so far one person's claim, but the way it can be checked deserves the attention of practitioners.
What Alpöge claimed
On Friday, Alpöge published a note on his personal site: according to him, Claude helped build the solution in about six weeks. He writes that he had worked with the model for several months before, but at the finish line it "far outpaced" the author himself. The note contains an almost poetic line saying Claude "flew, danced and sparkled" at the finish.
Bernhard Riemann formulated the problem in 1859, so it is nearly 170 years old. It concerns prime numbers, those divisible only by one and by themselves. Riemann asked how often primes occur and how they are distributed along the number line, and he derived a formula for estimating how many primes lie below a given bound. The hypothesis is on the Clay Institute's list of seven "Millennium Prize Problems," and the institute offers $1 million for solving each one.
The timing is awkward: the note appeared the day after Anthropic listed on Nasdaq. Alpöge himself acknowledged: "The timing is terrible. If I could, I would have finished sooner."
Why Lean matters here
Alpöge attached code written in Lean and argues that it allows the result to be checked. Lean is a formal verification system. A proof is written as a program, and the computer rigorously checks every logical step. This is similar to unit tests: either everything holds together or the check fails. That sidesteps the main weakness of language models, their tendency to write smooth but wrong reasoning.
Caveats remain. When code passes Lean successfully, that does not by itself guarantee it encodes the Riemann hypothesis itself rather than a simplified version. Matching the formulations is a job for humans, and mathematicians have already taken it on. At the time of publication, there is no confirmation.
Competitors and skepticism
Alpöge is not the only contender. According to The New York Times, on Friday independent Italian mathematician Alessandro Massullo posted an incomplete proof he had worked on for nearly 12 years. He did not use AI, because he does not fully trust such systems.
There is also the backdrop of lab competition. OpenAI previously said a closed version of its models helped solve part of another "Millennium Prize Problem." Many were surprised by the approach: a company worth nearly a trillion dollars spent $15 million and deployed 10,000 AI systems. Alpöge was working for Anthropic on that separate problem at the time.
Some mathematicians view such stories poorly. James Robinson of the University of Warwick said in September that it saddened him to see large companies "burning fuel" just to brag about a new model. He called it "immature playground boasting," backed by billions of dollars and a risk of climate harm.
What this means for practitioners
First takeaway: the pairing of "model proposes, formal verifier checks" is today the most reliable way to get AI results you can trust. The same principle works at smaller scales: tests for generated code, schemas for agent outputs, automated checks for critical business logic. The stricter the check, the less you need to take the model's word for it.
Second takeaway: until mathematicians confirm the proof, it cannot be treated as fact. If it survives expert scrutiny, it will be the strongest argument yet for AI as a research tool. If not, it stays a reminder that a loud claim and a verified result are different things, especially when a company has just completed a share offering.
- Nov 21, 2026OpenAI Opens ChatGPT Spaces to Plus Subscribers: Shared Documents with AI
- Nov 21, 2026Chinese Tech Giants Double AI Spending to $139B, While Labs Recover Less Than 8%
- Nov 21, 2026Ego lite Gives Each AI Agent Its Own Workspace Space in the Browser
- Nov 21, 2026Docker rebranded for AI agents, but Sandboxes shipped back in September
Key Questions Answered
What Alpöge claimed?
On Friday, Alpöge published a note on his personal site: according to him, Claude helped build the solution in about six weeks. He writes that he had worked with the model for several months before, but at the finish line it "far outpaced" the author himself. The note contains an almost poetic line saying Claude "flew, danced and sparkled" at the finish. Bernhard Riemann formulated the problem in 1859, so it is nearly 170 years old. It concerns prime numbers, those divisible only by one and by themselves.
Why Lean matters here?
Alpöge attached code written in Lean and argues that it allows the result to be checked. Lean is a formal verification system. A proof is written as a program, and the computer rigorously checks every logical step. This is similar to unit tests: either everything holds together or the check fails. That sidesteps the main weakness of language models, their tendency to write smooth but wrong reasoning. Caveats remain.
What are the key details on Competitors and skepticism?
Alpöge is not the only contender. According to The New York Times, on Friday independent Italian mathematician Alessandro Massullo posted an incomplete proof he had worked on for nearly 12 years. He did not use AI, because he does not fully trust such systems. There is also the backdrop of lab competition.
Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews
Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.
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.