explainx.ai350K+ learn from us
TrendingAI News TodayExplainx for Business
Pricing
explainx.ai

Upskill in AI — 16 free pathways, live workshops & bootcamps, and 50+ courses from practitioners. Plus the skills, tools, and MCP servers to practice on.

follow us

follow on google

Add explainx.ai as a preferred source

corporate training

support@explainx.ai

get started

Find your pathTake Free Evaluation

community

Join the community

learn

mind: share how you thinkpathways — start freeworkshopsbootcampscoursescompare Explainxcertificationsmock testsexplainx universitycorporate traininglearn skills & mcp

discover

skillsmcp serversexplainx mcpestacktoolsmdx readeragentsllmsdesignsdictionarypeopleagi trackerfelony benchranks

company

aboutvisionmissionteaminstructorsteach on explainxpartnershipscommunityhackathonscareers

content

daily AI newsstate of AI — live resultsblogreleasespromptsgeneratorsresource libraryfor LLMsexplainx.ai kids

solutions

all solutionsdeveloper upskillingmarketing upskillingproduct manager upskillingleadership upskilling

newsletter · weekly

Get AI news, tools, and insights in your inbox.

supportcontactprivacytermsdata rightshow we create contentsubmission guidelines

© 2026 AISOLO Technologies Pvt Ltd

explainx.ai

On this page

  • Update (November 20, 2026, 17:30 UTC): Verge and Willison amplify the same claim
  • TL;DR: the claim and its status
  • What exactly did Alpöge say?
  • What is the Riemann Hypothesis, in plain words?
  • What is verified today?
  • How does a Lean proof get checked?
  • What are developers and mathematicians saying?
  • How does this fit the 2026 pattern of AI math claims?
  • Why the Lean statement is the first thing to check
  • What should builders and founders do now?
  • What would confirmation mean?
  • How to follow the verification
  • Related reading
← Back to blog

explainx / blog

Levent Alpöge Claims a Claude-Assisted Proof of the Riemann Hypothesis: What Is Claimed, What Is Verified

Riemann Hypothesis, Anthropic, Claude, AI Mathematics, Lean

Part of Anthropic and Claude

Anthropic mathematician Levent Alpöge says he proved the Riemann Hypothesis with Claude and posted Lean code. What is claimed, what is verified, how to check.

Nov 20, 2026·10 min read·Yash Thakker
add explainx.ai
go deep
Levent Alpöge Claims a Claude-Assisted Proof of the Riemann Hypothesis: What Is Claimed, What Is Verified

Levent Alpöge, a mathematician who works at Anthropic, says he and Claude have found a proof of the Riemann Hypothesis, and of its generalized form. He posted the claim on November 20, 2026 on his personal website, with a PDF note and, according to Scientific American, Lean code that is meant to check each step. No independent expert has confirmed the proof yet. This post separates what is claimed from what anyone can verify today.

Update (November 20, 2026, 17:30 UTC): Verge and Willison amplify the same claim

The Verge published a mainstream-tech summary the same afternoon, framing Alpöge as an Anthropic mathematician claiming a Millennium Prize problem with Claude's help. The piece does not report independent verification; it contextualizes the claim beside earlier disputed AI-and-math headlines explainx.ai fact-checked in Navier-Stokes rumor posts.

Simon Willison's weblog reposted extended quotes from Alpöge's statement, including the line that formalization is finished while a readable write-up is still to come. Willison's item is discovery for developers, not peer review. The verification checklist below is unchanged: compile Lean, read the theorem statement, wait for experts.

Daily briefFrom Yash · 350K+ students taught

Catch up on AI

One AI story explained in plain English, plus one thing to try. Every day, free. Unsubscribe anytime.

TL;DR: the claim and its status

table · 2 cols
QuestionShort answer
Who claims it?Levent Alpöge, an Anthropic mathematician, on his own website
What is claimed?A proof of the Riemann Hypothesis and the Generalized Riemann Hypothesis, found with Claude
Is there machine-checkable evidence?Alpöge posted Lean code, per Scientific American
Has it been independently verified?No. Experts have not yet reported a successful check
Is it an Anthropic announcement?No company post found; it is a personal statement
What should you do?Wait for independent Lean builds and expert write-ups before treating it as settled

What exactly did Alpöge say?

Scientific American reporter Joseph Howlett quotes the statement directly: "I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis." Note the wording. It says "I believe", and it credits Claude as a collaborator, not only a tool.

The magazine adds that Alpöge says the result confirms the Generalized Riemann Hypothesis, a much larger statement that, if true, settles many open questions in number theory that currently assume it.

Several Hacker News readers quoted the note itself. One passage, as quoted in a comment by yewenjie, says the written argument is "incredibly long" and that its formalization is finished, while a readable version still has to be written. In the same thread, commenter elffjs reported that the full written argument is over 250,000 pages long and not provided. We could not confirm that page count from a second source, so treat it as one reader's reading of the note.

The note's tone is also notable. According to Scientific American, Alpöge writes that it is natural to feel "that this is an end" for a mathematician, while saying he believes it is "one dawn across science of something far greater."

What is the Riemann Hypothesis, in plain words?

Prime numbers (2, 3, 5, 7, 11 and so on) look scattered with no pattern. In 1859 Bernhard Riemann proposed that their distribution is tied to the zeros of a function called the zeta function, and that all the interesting zeros sit on one straight line. Scientific American describes this as a claim that the primes are "determined by an intricate, meaningful order."

For 167 years nobody could prove it. The magazine says most top experts spent little time on it because there was no obvious path. Many theorems in number theory today begin with "assume the Riemann Hypothesis". A proof would turn those conditional results into unconditional ones.

The generalized version covers a whole family of related functions, not just the zeta function. That is why the claim is so large: if both are true, a great deal of mathematics changes status overnight.

What is verified today?

Be strict here. Three things are verified as facts:

  1. The statement exists. Alpöge's site hosts the note, and Scientific American reports that he published it.
  2. Alpöge's background is real. His site lists his role at Anthropic, his Harvard fellowship and his earlier published work, including a solution of Hilbert's tenth problem over all number fields.
  3. A Lean file has been posted, per Scientific American. That is checkable by anyone with Lean and enough compute.

Three things are not verified:

  1. That the proof is correct. No expert report exists yet.
  2. That the Lean statements match the real Riemann Hypothesis. This is the most common failure in formalization. A proof can compile perfectly and still prove a slightly different statement.
  3. That Claude did most of the work. Alpöge credits Claude strongly, but the split between human and model is his own account.

How does a Lean proof get checked?

A chain of connected proof steps ending in a green check, showing how a Riemann Hypothesis proof is machine-checked in LeanA chain of connected proof steps ending in a green check, showing how a Riemann Hypothesis proof is machine-checked in Lean

Lean is a proof assistant. You write a theorem and a proof in a formal language. The Lean kernel then checks every step. If it compiles without errors, the logic is sound relative to the stated axioms.

Scientific American says "issues can arise in Lean," and that experts should be able to determine quickly whether the hypothesis is coded correctly and whether the proof holds. In practice, a reviewer will do four checks:

  • Read the main theorem statement. Does it say what the Riemann Hypothesis says? Are the definitions of the zeta function, its zeros and the critical line the standard ones, or custom ones that quietly change the meaning?
  • Look for sorry and added axioms. In Lean, sorry marks a missing proof. A custom axiom assumes a fact without proving it. Either one breaks the claim.
  • Build the whole project from a clean machine. A proof that only compiles on the author's setup is not enough.
  • Inspect the dependencies. The proof should rest on the standard mathematics library, not on unreviewed local files.

You can learn more about this style of checking from our earlier coverage of Lean validation of an OpenAI Yang-Mills result and the claimed-vs-verified analysis of quantum-problem Lean proofs.

What are developers and mathematicians saying?

The Hacker News thread had about 250 points when we read it. Reactions split in three ways. These are commenters' views, not verified facts.

Trust in the author. User gus_massa wrote that this is not a random employee, and that Alpöge already has several verified results this year. User xianshou called him one of the best in the world at guiding AI toward new mathematics. Another commenter, jrflo, said that because Lean code is included and the author has a reputation, the cautious qualifiers in the post title could be dropped. That is an opinion. Compilation by independent parties is the real test.

Questions about release. User yewenjie first asked why the formalization was not released while the human-readable write-up is pending, then edited the comment to say the Lean formalization is linked. Check Alpöge's site for exactly which files are public.

Worries about understandability. User puff_pastry worried that a result nobody can read in full moves away from science where a person can follow the steps. User ariwilson liked that the claim came from a human mathematician rather than a marketing post, while feeling sad that the credit looked like a 60/40 split toward AI. Others raised the timing: the claim arrived a day after Anthropic's IPO. That is an observation by commenters such as albinstman, not evidence of anything.

How does this fit the 2026 pattern of AI math claims?

AI mathematics claims have come in a rapid stream this year, and several turned out to need careful reading. This post is not a judgment on any of them, but the history is useful:

  • July: Alpöge tweeted a Claude-generated result on the Jacobian conjecture. See our coverage of the Fable 5 Jacobian counterexample and what the Jacobian conjecture is.
  • August: A Claude result on a Riemann zeta lower bound showed partial progress on related territory.
  • October: OpenAI published a large batch of manuscripts. See what to check in the 722 manuscripts and expert reactions to OpenAI's math results. Scientific American says one of those proofs made significant progress on understanding the primes.
  • Other labs: Google, Meta and DeepMind each reported open-problem results, such as Google's five math problems, Meta Muse Spark's six open problems and AlphaProof Nexus.

The lesson from that run: results with a short, machine-checked statement have held up best. Results that rely on long text and trust have taken longer to settle. Alpöge's claim has a Lean artifact, which is the stronger form, but its size makes review harder.

Why the Lean statement is the first thing to check

A magnifier over a claim marked with a green check, illustrating checking a Riemann Hypothesis claim before believing itA magnifier over a claim marked with a green check, illustrating checking a Riemann Hypothesis claim before believing it

When a formal proof of a famous problem appears, the first doubt is rarely "is the logic wrong?" Lean makes sure of that. The first doubt is "did the author prove the right statement?"

Common ways that goes wrong in formalization work:

  • A definition is simplified, for example restricting the zeta function to a smaller domain.
  • A hypothesis is added that the original problem does not have.
  • The theorem is stated about a related but weaker object.
  • A tactic or library bug lets an incorrect step through (rare, but it happens).

Mathematicians on the Lean community channels can usually spot these in hours, not weeks, because the main theorem statement is short even when the proof is huge. That is why Scientific American says verification "shouldn't take long." It is also why you should wait for a named, independent party to say "the statement is right and it builds."

What should builders and founders do now?

You do not need to act on this today. A few practical points:

  • Do not repeat the headline as fact. "Anthropic proves Riemann Hypothesis" is wrong on two counts: it is not confirmed, and it is not a company announcement. The accurate form is "an Anthropic mathematician claims."
  • Watch for independent builds. The most reliable signal is someone outside Anthropic posting that the Lean project compiles and that the main statement is the standard one.
  • Watch for the human-readable paper. Alpöge says he intends to write it properly. Number theorists will judge the ideas from that.
  • Use it as a case study in verification. Machine-checkable output changes how you trust AI-generated work. If you build agents that produce code or proofs, the same principle applies: produce an artifact a separate checker can run. Our workshops cover this style of agent evaluation.

If you follow AI claims in math, our earlier piece on labs holding back math solutions is useful background for why disclosure and checkability matter.

What would confirmation mean?

Scientific American says that if the proof stands, it would cap a year in which AI's mathematical ability went from "unreliable college student to a superhuman force." That is the magazine's framing, and it is a strong one. A more careful reading is narrower: it would show that a human expert working with a frontier model can produce a formally checked proof of a problem that resisted experts for 167 years.

For Anthropic, this is a claim made by an employee, and no company statement is known. The IPO-timing comments on Hacker News are speculation. The facts to hold on to are the Lean artifact and the reviews that follow.

For the field, a verified result would change the status of a large set of conditional theorems. It would also raise a new problem: how to teach, read and audit proofs that a person cannot read end to end. Alpöge's own note, as quoted by commenters, shows he is aware of that, since he says he plans to write a readable version.

How to follow the verification

Check these places for updates:

  • Alpöge's website, for new files, corrections or the readable paper.
  • The Scientific American article, which the magazine says will be updated.
  • The Hacker News discussion, where people may post their own Lean build results.
  • The Lean community's public channels, where formalization experts usually report build results first.

We will add dated updates to this post when independent checks appear.

Related reading

  • Fable 5 Jacobian conjecture counterexample and Alpöge
  • Claude Riemann zeta lower bound (August 2026)
  • OpenAI's 722 math manuscripts: what to check
  • Expert reactions to OpenAI's math results
  • AlphaProof Nexus and peer-reviewed Erdős problems
  • Anthropic IPO priced at $83

Details are accurate as of November 20, 2026. The claim is new and unverified, and facts may change as experts review it.

Spotted something out of date? Let us know.
Daily briefFrom Yash · 350K+ students taught

Catch up on AI

One AI story explained in plain English, plus one thing to try. Every day, free. Unsubscribe anytime.

  • ✓ Free
  • ✓ One email a day, no spam
  • ✓ Unsubscribe anytime

People in this article

  • Simon Willison →Independent open source developer and creator of Datasette
Explore people in AI →
Yash Thakker

Written by

Yash Thakker

Yash is an AI expert with over 300K learners. Join his workshops →

View Yash Thakker in People in AI →

Related posts

Sep 23, 2026

Boris Cherny Used Opus 5.5 to Formally Verify the Claude Agent SDK

A couple of short prompts turned into 16 pull requests fixing race conditions and state-management bugs Boris Cherny says a human likely wouldn't have spotted. He used Claude Opus 5.5 to formally model the Claude Agent SDK in Lean 4 and TLA+ — 1,529 theorems, zero unproven "sorry" gaps, 19 of 24 bugs found directly by the proofs. Here's what formal verification by an agent actually looks like in practice.

Sep 5, 2026

Claude Wrote the First Machine-Checked Proof of Fermat's Last Theorem

On September 5, 2026, Anthropic announced that Claude completed the first fully formalized, machine-checked Lean 4 proof of Fermat's Last Theorem — a project mathematicians expected to take years, done in 11 days and totaling more than 13 million lines of code, the largest Lean proof ever written.

Nov 20, 2026

Claude Chat Flagged, FBI at the Door: What Is Claimed and What Is Verified

A researcher published one user's account of an FBI and police visit after Anthropic banned his Claude account and surfaced a chat. The author cannot verify it. We separate the claims from what Anthropic's policies and earlier reported cases confirm.