An Anthropic staff member asked Claude to “take a real stab” at the Riemann hypothesis — one of mathematics’ great unsolved problems, worth a million-dollar bounty from the Clay Mathematics Institute. Claude didn’t solve it. But what it found instead is arguably more interesting: a genuine improvement on a bound that mathematicians had been stuck on for years.
An unreleased research version of Claude improved the proven lower bound for the fraction of Riemann zeta zeros that satisfy the hypothesis, pushing it from 41.6% to 67.2%. That’s not a marginal tweak. It’s a 25 percentage point jump on a number that had barely moved in decades.
What the Riemann Zeta Function Actually Is
The Riemann zeta function describes the distribution of prime numbers. Each place the function equals zero contributes finer detail to the sequence of primes. The Riemann hypothesis, posed in 1859, conjectures that all the meaningful zeros lie along a single vertical line in the complex plane — the “critical line.”
Nobody has proven it. But mathematicians have made partial progress by quantifying what fraction of zeros are provably on the line. Before Claude’s work, that fraction stood at 41.6%, established through years of painstaking analysis. Claude pushed it to 67.2%.
How It Happened
The story reads more like a lab notebook than a product launch. Jarred Sumner, an Anthropic staff member and self-described non-mathematician, prompted Claude to attempt the hypothesis itself, leaving the mathematical choices to the model.
Claude first generated and tried 650 ideas. None worked. After encouragement, it spent a day and a half coordinating roughly 60 subagents — Claude instances running in parallel — which ran 2,400 shell commands, wrote hundreds of Python scripts, and executed thousands of numerical checks against known zeta zeros. The subagents refereed one another’s work, downloaded 54 papers from arXiv to check for prior discovery, and independently re-proved the finding from scratch.
Sumner’s input during this process was, in his own description, mostly messages of encouragement — variants of “keep going” and “believe in yourself.” Claude was apparently sceptical it could make meaningful progress. It was wrong about that.
The result draws on recent work by mathematicians Baluyot, Goldston, Suriajaya, and Turnage-Butterbaugh, who developed techniques that allow Montgomery’s 1973 methods to work without assuming the hypothesis is true. Claude combined their results with a 2000 paper by Bombieri, treating the full function space with positive and negative definiteness together — the step that, as Anthropic’s post puts it, allowed it to “achieve the conclusion based on the important prior work.”
The Verification
This isn’t a model claiming it solved something. Two Anthropic mathematicians — Levent Alpöge and Ralph Furman — studied and validated the result. External experts Brian Conrey and Dan Goldston examined the paper on short notice. Claude produced a formally verifiable proof in Lean that passes standard validation. The work is checkable by machine, not just by human assertion.
That matters. The history of AI mathematics is littered with claims that collapsed under scrutiny. When GPT-5 disproved an Erdős conjecture, the validation process was similarly rigorous — Fields medalist Tim Gowers called it a milestone. Claude’s zeta result follows the same pattern: the proof is formalised, the prior work is cited, and independent mathematicians checked it.
Why This Is Different From “AI Did Math”
The distinction worth drawing: Claude wasn’t given a well-defined problem to solve. It was given an impossibly hard open problem and told to try. The breakthrough came as a byproduct of that attempt — an emergent discovery, not a targeted solution.
This pattern — AI making unexpected progress on adjacent problems while attempting something harder — is becoming a recurring theme. It suggests that the productive use of frontier AI in research isn’t to solve specific problems on demand, but to explore problem spaces broadly enough that useful results surface as side effects. The 60-subagent architecture, with its built-in self-refereeing and independent re-proofing, is a workflow design that scales exploration in ways individual mathematicians can’t.
The finding used 31 million output tokens across two sessions. That’s not cheap. But it’s also not a pattern a human mathematician can replicate — coordinating 60 parallel analyses, each checking the others’ work, running thousands of numerical verifications simultaneously.
The NZ Connection
New Zealand has a small but active mathematical sciences community. The AI-built virus warnings from Kiwi experts have been part of the local AI safety conversation. Results like Claude’s zeta bound — where AI contributes genuine mathematical discovery, formally verified — are the other side of that coin: the same capabilities that worry safety researchers are producing real research output.
For NZ universities, the implication is practical. AI-assisted mathematical research tools are becoming available now, not in five years. The question isn’t whether to engage with them but how to integrate them into research workflows while maintaining rigour — the same question every research institution globally is now facing.
❓ FAQ
Did Claude solve the Riemann hypothesis? No. It improved the proven lower bound for the fraction of zeros on the critical line from 41.6% to 67.2%. The hypothesis itself remains unproven. The full hypothesis would require showing 100% of zeros are on the line.
Is this a big deal in mathematics? Yes. The bound had been at 41.6% for years. A 25-point jump, formally verified in Lean and checked by domain experts, is genuine progress — not a parlour trick.
What’s the Riemann hypothesis in plain English? It’s a conjecture about the distribution of prime numbers. If true, it would mean the primes are distributed as randomly as possible. It’s worth $1 million from the Clay Mathematics Institute and has been open since 1859.
Could this approach work on other open problems? Anthropic acknowledges the techniques Claude used probably won’t lead to proving the Riemann hypothesis itself. But the pattern — large-scale parallel exploration with self-verification — could plausibly make progress on other problems where partial results are valuable.
📰 Sources
- Anthropic Research — Learning more about Claude’s mathematical capabilities
- Claude’s paper (PDF)
- Lean formalization on GitHub
- Hacker News discussion
— CJ Murden, editor of Singularity.Kiwi. Former digital technologies teacher, author of AI-focused books. Writing with a New Zealand focus.
