Anthropic mathematician Levent Alpöge says he and Claude have found a proof of the Riemann hypothesis, a conjecture that has stood since Bernhard Riemann proposed it in 1859. He posted the claim on Friday, November 20, 2026, in a four-page note on his website, along with a Lean code archive that outside experts have not yet fully checked. Scientific American framed the claim around what it called the most famous problem in all of mathematics.
The claim is big, and the public record is still thin: a note, a code archive and a few early reactions. Verification is the story now, and it will move quickly over the next few days as reviewers work through the files.
What Alpöge says he proved
Alpöge's central statement is short. "I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis." According to the note, the method also covers the generalized Riemann hypothesis, and the entire proof is already written in Lean, a proof assistant. Pasquale Pillitteri summarized the same note.
The scope needs a careful read. A CASRAI summary describes the release as an assumption-free formalization covering both the classical Riemann hypothesis and the generalized version for all Dirichlet L-functions. Kingy.ai reports that the extended Riemann hypothesis is future work rather than proved. Those accounts are hard to reconcile, so readers should check which version of the claim a given report means before repeating it.
Early reaction has been strong. A mathematician named Sutherland said he had independently compiled Alpöge's result in Lean, and he said, "This is indeed an historic moment." The coverage reviewed for this story does not say who Sutherland is or where he works. Treat that remark as the reaction of someone who ran the code, not as a formal expert review.
Why the Lean archive is so large
The release comes in 14 compressed parts. Together they total 239.3 megabytes, but they expand to more than 61.5 gigabytes of proof data. One outlet that streamed the files counted 239,275,554 bytes in the joined archive and 175,187 Lean source files, according to Kingy.ai.
| Item | Reported figure |
|---|---|
| Announcement | Friday, November 20, 2026 |
| Archive parts | 14 |
| Compressed size | 239.3 MB |
| Expanded size | More than 61.5 GB |
| Lean source files | 175,187 |
| Note length | Four pages |
Numbers like these explain the verification problem. No human can read 175,187 files line by line in a few days. Reviewers will lean on automated tools, and then they must trust those tools too. The note itself challenges universities and math departments to rethink peer review for very large AI-generated formal codebases.
Claude's role is also contested. Alpöge says Claude assembled a timeline of the work after searching through it, and he credits Claude with major technical advances once he pointed the project toward a promising approach, as Superpower Daily reported. The note reportedly describes that timeline as rough and warns that some descriptions may need correction. That caveat matters, because it means the public account of who contributed what is not final.
What Lean can and cannot confirm
Lean checks whether each step of an argument follows from formal logical rules. That is a demanding test, and it is one human referees rarely apply to a proof this long. A proof checker answers a narrower question than the one mathematicians care about, though. It confirms that the argument is valid for the statement as written. It does not confirm that the statement is the Riemann hypothesis the field has in mind.
That is why reviewers must inspect how the statement is encoded, along with the definitions and assumptions the proof depends on. A proof can compile cleanly and still establish something slightly different from the conjecture. Experts have to check the setup as well as the steps.
Has the Riemann hypothesis been solved?
Not by any standard the field has accepted. The conjecture says that the nontrivial zeros of the Riemann zeta function all lie on the critical line, the vertical line where the real part equals one half. The Clay Mathematics Institute lists the Riemann hypothesis among its seven Millennium Prize Problems. Its prize rules require a proof to appear in a refereed journal of recognized standing and to hold up for two years afterward before an award is considered.
Those rules show how high the bar is. A claim like Alpöge's will be judged by many readers over a long stretch, not settled in a single weekend. The Interesting Engineering headline frames the claim as an attempt to crack a 167-year-old problem, and that framing is accurate about the age of the question. It says nothing yet about whether the answer is right.
Where verification stands
The reports disagree on the most basic point. One account says the archive was verified by MIT kernel checks. Another says only the Lean code is public and no outside expert has checked it yet. I did not find an independent expert review in the coverage reviewed for this story. Anthropic has not formally endorsed the claim in the material reviewed either.
Four questions will decide how the story ends:
- Does the Lean statement faithfully encode the Riemann hypothesis?
- How much of the work came from Claude, and how much from Alpöge?
- Does the generalized version hold, or is it still future work?
- Will Anthropic back the claim on the record?
How this fits Anthropic's broader record
The story has an earlier Claude-linked thread. A 2026 arXiv preprint describes a Claude-produced proof that more than two-thirds of the nontrivial zeros of the zeta function lie on the critical line. The preprint uses that work as a test case for measuring how non-trivial machine-generated proofs are. I did not verify that result, so treat it as the preprint's own description rather than settled mathematics.
The timing also carries weight. CNBC reported Anthropic's trading debut with a $1.8 trillion market cap. Public-market investors now have a stake in whether research claims like this one hold up, and USA 24 has covered the debut and its opening trade in detail. A verified proof would be a major capability story. A failed check would damage credibility.
Outside scrutiny of Anthropic's capability claims is not new. A METR analysis said Anthropic's AI improves twice as fast, but most of its evidence was redacted. That is a useful reminder that claims about AI progress depend on what outsiders can actually inspect.
What readers should do now
- Treat the proof as unconfirmed until independent mathematicians or a recognized journal or preprint review confirms it.
- Check which version of the Riemann hypothesis a report describes, classical or generalized, before you repeat it.
- Do not trade on the headline. Anthropic's stock is newly public, and a single claim should not drive a position.
- If you use AI tools for research, check their outputs against established references and primary sources.
- For Claude users and enterprise buyers, this is a capability claim, not a product change. It does not, by itself, alter pricing or access.
What to watch in the coming days
Watch for detailed reviews of the Lean statement and proof steps, not just reports that the code compiles. Watch whether Alpöge posts the work as a preprint or submits it to a journal, and how quickly referees respond. Also watch for an official Anthropic statement, any revised version of the note or archive, and whether universities adopt new review rules for very large AI-generated codebases.
The next few days will test whether a 61.5 GB formal archive can be audited quickly enough to matter. If the statement holds up, the result would close a question that has waited 167 years. If it does not, the episode will still show how fast an unverified proof can travel.




