Exploring the Mathematical Prowess of Claude

On August 10, 2026, Anthropic shared a fascinating account of an "unreasonable challenge" posed to Claude by a staff member. The goal was audacious: attempt to solve the Riemann hypothesis, one of the most notorious unsolved puzzles in the history of mathematics.

While Claude did not crack the hypothesis—a feat that has eluded humans since 1859 and carries a million-dollar prize reward of $1,000,000—it achieved something unexpected. In the process of its failure, it made a significant breakthrough on a closely related mathematical problem.

The Breakthrough: Improving the Lower Bound

An experimental, unreleased version of Claude managed to enhance a long-standing lower bound regarding the proportion of zeros of the Riemann zeta function that align with the Riemann hypothesis.

Comparison of Results

This leap was made possible by synthesizing decades of existing mathematical research. The result was meticulously reviewed and validated by two mathematicians at Anthropic, who then drafted a concise technical note for the expert community. Furthermore, Claude provided a formally verifiable proof.

"We are grateful to Brian Conrey and Dan Goldston, two experts in this area, who generously examined the paper on short notice."

While Anthropic notes that these specific methods likely won't solve the Riemann hypothesis itself, the event underscores the accelerating trajectory of AI's mathematical capabilities.

---

Understanding the Mathematics

The Riemann Zeta Function

The $\zeta(s)$ function is critical because it describes how prime numbers are distributed. The "zeros" of this function provide the fine-grained detail of the sequence of primes.

• The Hypothesis: All non-trivial zeros lie on a single vertical line (the critical line).
• The Impact: If proven, it would provide a foundation for the "randomness" observed in prime distributions.

The Research Lineage

Claude didn't work in a vacuum; it built upon a specific lineage of analytic number theory:

1. Montgomery (1973): Introduced techniques for zero distribution (though these originally assumed the hypothesis was true).
2. Modern Era: Researchers including Baluyot, Goldston, Suriajaya, and Turnage-Butterbaugh adapted Montgomery's work to function without that assumption.
3. Bombieri (2000): Provided further foundational work.

Technical Deep Dive

Claude's approach involved the following logic:

• It constructed a specific space of functions.
• It utilized a quadratic form induced by Weil.
• It identified positive-definite and negative-definite subspaces based on whether zeros were on or off the critical line.
• It applied an inequality regarding the rank of a quadratic form using first- and second-moment data.

The key innovation was Claude's willingness to treat the entire space as a whole—allowing the quadratic form to be non-diagonal and accounting for both positive and negative definiteness simultaneously.

---

The Methodology: How Claude Did It

The discovery occurred within an environment called Claude Code over two distinct sessions, consuming approximately 31 million output tokens.

The Workflow

The process was initiated by Jarred Sumner, a non-mathematician, who gave Claude a simple, open-ended prompt:

"Take a real stab at the Riemann hypothesis."

The Human Element

Interestingly, Jarred's role was not technical. He primarily provided emotional support, using phrases like:

• "Keep going!"
• "Believe in yourself!"

This encouragement reportedly helped the model push past its own internal skepticism regarding the possibility of making a breakthrough.

Rigorous Self-Verification

Before presenting the result, Claude performed its own quality assurance:

• Deployed subagents to review the proofs.
• Searched for potential counterexamples.
• Downloaded 54 papers from arXiv to ensure the result was original.
• Independently re-derived the proof from scratch.

Final Validation

Once the result was found, Claude suggested human verification. Levent Alpöge and Ralph Furman (Anthropic mathematicians) analyzed the work to confirm its validity and relationship to prior research. Simultaneously, Claude collaborated with Eric Easley to translate the proof into Lean, a formal verification language.

-- Conceptual representation of a formal proof step
theorem riemann_zeta_lower_bound : 
  proportion_of_zeros_on_line ≥ 0.672 := 
begin 
  -- Claude's proof logic here
end