Here is the part that sounds like a joke but is not: the human who prompted Claude is not a mathematician. Jarred Sumner, an Anthropic staff member, told the model to "take a real stab" at the hypothesis, then left the mathematical choices to the model . When Claude's first 650 ideas all failed, Sumner's main contribution was sending messages of encouragement — "keep going," "believe in yourself".
Claude kept going. Over the next day and a half, it coordinated roughly 60 sub-agents that executed 2,400 shell commands, wrote hundreds of Python scripts, ran thousands of numerical checks against known zeta zeros, and refereed one another's work . The result: a 25.6-point jump that venture capitalist Deedy called "the most significant advance in analytic number theory since the bounded prime gap result in 2013".
The technical approach is as surprising as the result. Claude built on two lines of prior work: Hugh Montgomery's 1973 techniques (which assumed the hypothesis was true) and recent papers by Baluyot, Goldston, Suriajaya, and Turnage-Butterbaugh that removed that assumption . The model combined those with a 2000 paper by Enrico Bombieri. But the key insight was methodological: Claude treated the entire function space together — positive- and negative-definite subspaces, non-diagonal quadratic forms — rather than separating them as mathematicians had traditionally done .
That sounds abstract. What it means is: Claude found a path that human mathematicians had not thought to take.
Anthropic's own mathematicians, Levent Alpöge and Ralph Furman, validated the result. Two external experts — Brian Conrey and Dan Goldston — reviewed it on short notice . Claude also generated a Lean formal proof, pushing the result beyond "paper accepted" into "machine-verified" territory .
The result does not prove the Riemann hypothesis. A 67.2% lower bound is not 100%, and the gap is not a matter of "finishing the remaining 32.8%." But the deeper signal is not the number. It is the workflow: 60 sub-agents, autonomous coordination, 31 million tokens, and a non-expert human providing emotional support.
This is the second time in a week that AI has touched an open research problem in mathematics . It follows the GPT-5.6 and Fable 5 collaboration on MIMO detection. And it comes as a new generation of talent is moving into AI labs — John Jumper from Google DeepMind, Jacob Tsimerman from Toronto . The message from Anthropic's CEO Dario Amodei was direct: don't think of AI as a tool for analyzing data. Think of it as an agent that can do the work of a scientist end to end .
The Riemann hypothesis is still unsolved. But the way the gap was closed — by 60 autonomous agents running on a model that hasn't even been released — says more about where research is heading than any single number ever could.
P.S. Anthropic emphasized one more thing: "We don't expect that the techniques Claude used will lead to proving the Riemann hypothesis" . That is a careful disclaimer from a company that knows exactly what it just demonstrated. The result isn't the path to the proof. The method is the signal. And the signal is loud.
Frequently Asked Questions
Q: Did Claude actually prove the Riemann hypothesis?
A: No. The Riemann hypothesis, first posed in 1859, remains unsolved. Claude was asked to "take a real stab" at it and did not succeed. What it did instead was improve a related mathematical result: the proven lower bound for the fraction of Riemann zeta function zeros that lie on the critical line .
Q: What exactly did Claude achieve?
A: Claude raised the proven lower bound from 41.6% to 67.2% — a 25.6 percentage point jump. For context, in the previous 37 years, mathematicians had advanced that same bound by just 0.8 percentage points . Menlo Ventures partner Deedy called it "the most significant advance in analytic number theory since the bounded prime gap result in 2013" .
Q: How did Claude actually do it?
A: Over two sessions in Claude Code, consuming 31 million output tokens, Claude coordinated about 60 sub-agents that executed 2,400 shell commands, wrote hundreds of Python scripts, and ran thousands of numerical checks . The process was not smooth: Claude first generated and tested 650 ideas, all of which failed. After being prompted to try again, it reorganized the sub-agents and found the breakthrough .
Q: Who prompted Claude, and what was their background?
A: Jarred Sumner, an Anthropic staff member and founder of Bun, prompted the model. He is not a mathematician. His main contribution during the process was sending encouragement like "keep going" and "believe in yourself" after Claude's 650 failed attempts .
Q: How was the result verified?
A: The result was validated through multiple channels. Anthropic mathematicians Levent Alpöge and Ralph Furman examined and validated the paper. External number theorists Brian Conrey and Dan Goldston — one of whom co-authored the prior work Claude's argument builds on — reviewed it on short notice. Claude also produced a Lean formal proof that passes the standard validation tool .
Q: Is the Lean formal proof a guarantee that the result is correct?
A: Not entirely. Lean checks that a formal theorem follows from its stated definitions and accepted axioms — but it does not certify that the formal statement faithfully encodes the informal claim in the paper. If the formalization quietly weakens a hypothesis, Lean will happily verify the weaker thing . Having Alpöge, Furman, Conrey, and Goldston restate and review the proof informally is what partially closes that gap .
Q: Why does the 67.2% number matter?
A: 67.2% is the proven minimum proportion of zeros that must lie on the critical line. The previous record was 41.6%, established by Conrey in 1989. It does not mean the Riemann hypothesis is "67.2% solved" — a full proof requires showing all zeros are on the line, and the gap from 67.2% to 100% is not a matter of "remaining 32.8%" . But the single-jump magnitude — 25.6 percentage points after decades of incremental progress — is what makes the result extraordinary .
Q: What was Claude's key technical insight?
A: Claude built on work by Hugh Montgomery (1973) and recent papers by Baluyot, Goldston, Suriajaya, and Turnage-Butterbaugh that removed the assumption that the hypothesis was true. It combined those with a 2000 paper by Bombieri. The methodological breakthrough was treating the entire function space — positive and negative subspaces together, allowing the quadratic form to be non-diagonal — rather than separating them as mathematicians had traditionally done .
Q: Can the techniques Claude used lead to a full proof of the Riemann hypothesis?
A: Anthropic explicitly says no. The company stated: "We don't expect that the techniques Claude used will lead to proving the Riemann hypothesis." The significance is not the path to the proof — it is the method: 60 autonomous agents, self-verification, cross-referencing 54 arXiv papers, producing Lean formal proofs, and writing a paper .
Q: Is this an isolated event, or part of a broader trend?
A: This is the second time in a week that AI has touched an open research problem in mathematics, following the GPT-5.6 and Fable 5 collaboration on MIMO detection. It also follows Anthropic's Fable 5 producing a counterexample to the 87-year-old Jacobian conjecture. The pattern across these events: models are not replacing mathematicians — they are extending the reach of existing human results by finding combinations nobody had assembled .
