Claude’s Riemann Hypothesis Attempt: A Significant Partial Result, Not a Solution
A claim that needs precision
A recent Anthropic report describes an unreleased research version of Claude making unexpected progress during an attempt to address the Riemann Hypothesis, one of mathematics’ most famous open questions [1]. The result is striking, but the headline must be stated accurately: Claude did not prove the Riemann Hypothesis, and it was not shown to be close to proving it. Instead, the system produced a new lower bound for the proportion of non-trivial zeros of the Riemann zeta function known to lie on the critical line.
Anthropic says that the bound rose from 41.6% to 67.2%. That is a meaningful advance on a related question in analytic number theory. It does not mean that 67.2% of the conjecture has been solved, nor does it establish that the remaining zeros lie on the same line.
“We don’t expect that the techniques Claude used will lead to proving the Riemann hypothesis.” — Anthropic [1]
That qualification is central. The achievement concerns a rigorously formulated partial theorem, subject to human scrutiny and formal verification, rather than a proof of the conjecture itself.
What the Riemann Hypothesis says
The Riemann zeta function, usually written as ζ(s), is a complex-valued function deeply connected with prime numbers. Its non-trivial zeros occur in a region called the critical strip, where the real part of the complex variable lies between 0 and 1. The Riemann Hypothesis proposes that every one of these zeros has real part exactly 1/2 [2].
In geometric terms, the conjecture says that all non-trivial zeros sit on one vertical line in the complex plane. The pattern matters because the location of those zeros controls how regularly prime numbers are distributed. The Clay Mathematics Institute lists the problem as unsolved and notes that many initial zeros have been checked computationally, but finite verification cannot replace a proof covering infinitely many cases [2].
Question | What is known from the reported work? |
Has the Riemann Hypothesis been proved? | No. Anthropic explicitly says Claude did not solve it. |
What improved? | A lower bound for the proportion of zeta zeros on the critical line. |
Reported change | From 41.6% to 67.2%. |
What does 67.2% mean? | At least that proportion can be established by the relevant theorem; it is not a percentage of the conjecture “completed.” |
Is the result machine-checked? | Anthropic reports a Lean formalization that passes the comparator validation tool, alongside human mathematical review. |
How the result emerged
The experiment began with a deliberately broad instruction from Jarred Sumner, an Anthropic staff member who is not a mathematician: ask Claude to “take a real stab” at the hypothesis. According to Anthropic, the first phase generated roughly 650 unsuccessful ideas. A second phase ran for about a day and a half and coordinated approximately 60 Claude subagents.
Across the two sessions, the system produced 31 million output tokens, executed about 2,400 shell commands, and wrote hundreds of Python scripts. The agents checked numerical cases against known zeta zeros, reviewed one another’s arguments, searched for counterexamples, consulted 54 papers from arXiv, and attempted an independent re-proof [1]. This is less like a single chatbot answer and more like a large, automated research workflow in which proposal, calculation, criticism, and revision were repeated many times.
The method also illustrates why raw model fluency is not enough for serious mathematics. A plausible-looking argument can fail because of a hidden assumption, a missing convergence condition, or an unjustified exchange of limits. The reported workflow therefore used separate agents as hostile reviewers and validators rather than treating the first coherent proof sketch as reliable.
What Claude actually contributed
Anthropic attributes the improved bound to a combination of existing research. The account names work by Aryan; by Baluyot, Goldston, Suriajaya, and Turnage-Butterbaugh; and by Bombieri [1]. Claude’s contribution was to identify a way to combine these ingredients through a suitable function space and a quadratic form associated with Weil’s framework.
In the concise description provided by Anthropic, positive- and negative-definite subspaces encode information associated with zeros on and off the critical line. An inequality involving the rank of a quadratic form is then paired with first- and second-moment estimates. The non-diagonal treatment of the quadratic form is presented as the step that permits the stronger lower bound.
The paper is technical, so the headline number should not be detached from its mathematical setting. A lower-bound theorem says that a specified fraction is guaranteed to lie on the critical line. It does not identify every zero, exclude all counterexamples, or imply that the unresolved part is merely an engineering problem.
Human review and formal verification
Anthropic reports that its mathematicians Levent Alpöge and Ralph Furman studied the result and prepared an informal note for experts. It also says that Brian Conrey and Dan Goldston examined the paper on short notice [1]. These reviews are important, but they should not be confused with the complete community process normally associated with a major mathematical proof.
Claude also helped produce a formalization in Lean, an interactive theorem-proving environment. The code is publicly available in Anthropic’s repository [4]. Formal verification can provide strong assurance that a formal statement follows from its encoded premises. It does not automatically prove that the encoded statement captures the most important interpretation of the informal mathematics, or that the theorem settles the original conjecture.
The distinction can be summarized simply: formal verification strengthens confidence in the reported partial result; it does not turn that partial result into a proof of the Riemann Hypothesis.
Why the result matters for AI-assisted mathematics
The most important implication may concern research practice rather than the conjecture’s final status. Claude was able to search a broad literature, generate many candidate approaches, run computations, organize specialized subagents, and return to a problem after repeated failure. That combination of language reasoning, code execution, tool use, and self-critique is different from asking a model to produce a proof in one response.
The experiment also suggests a division of labor. Human mathematicians supplied the standards of relevance and validation, while the model explored combinations at a scale that would be difficult for one person to sustain. The result therefore supports a view of AI as a research partner whose proposals require expert checking, not as an autonomous authority.
There are limits as well. The system’s success depended heavily on decades of published mathematics. Its output required expert interpretation. Its numerical experiments could test examples but could not establish an infinite claim. And the reported result remains a contribution to a related theorem, not a resolution of the central conjecture.
References
Anthropic: Learning more about Claude’s mathematical capabilities: https://www.anthropic.com/research/riemann-zeta
Clay Mathematics Institute: Riemann Hypothesis: https://www.claymath.org/millennium/riemann-hypothesis/
Clay Mathematics Institute: The Millennium Prize Problems: https://www.claymath.org/millennium-problems/
Anthropic: zeta-23-lean: https://github.com/anthropics/zeta-23-lean
Clay Mathematics Institute: Jeffery Vaaler, Riemann hypothesis: https://www.claymath.org/lectures/riemann-hypothesis/
From First Principles: What Claude Actually Did to the Riemann Hypothesis: https://www.youtube.com/watch?v=C-ECTnM7tUY
Comments