An unreleased research version of Anthropic's Claude has improved a longstanding mathematical result tied to the Riemann hypothesis, raising the known lower bound for the fraction of conforming zeros from 41.6% to 67.2%. The breakthrough, published by Anthropic on August 10, 2026, came about not as a targeted research project but as the unexpected byproduct of a casual challenge: take a real stab at one of the most famous unsolved problems in all of mathematics. For the latest developments in AI-driven discovery, follow AI Buzz Wire for ongoing coverage.
The Riemann hypothesis, first proposed by Bernhard Riemann in 1859, concerns the distribution of prime numbers and carries a one-million-dollar bounty from the Clay Mathematics Institute. It remains unproven. While Claude did not resolve the hypothesis itself, the model's work on a related sub-problem represents a genuine advance in number theory — one that human mathematicians have validated.
How the Challenge Began
The effort started when Jarred Sumner, an Anthropic staff member who is not a mathematician, prompted Claude to "take a real stab" at the Riemann hypothesis, leaving all mathematical decisions up to the model. Claude's initial attempt involved generating and testing approximately 650 ideas, none of which yielded results. After Sumner encouraged the model to try again, Claude spent roughly a day and a half coordinating approximately 60 subagents that explored the problem more deeply.
It was during this second attempt that Claude discovered a way to surpass the previous state-of-the-art lower bound. The model found that combining results from mathematicians Baluyot, Goldston, Suriajaya, and Turnage-Butterbaugh with earlier work by Bombieri provided a path to push the known proportion of zeros satisfying the hypothesis from 41.6% to 67.2%.
A Self-Checking, Self-Doubting System
What makes the result particularly striking is how rigorously Claude verified its own finding before presenting it. The model dispatched subagents to review the proofs, search for counterexamples, download 54 papers from the arXiv to confirm the result had not already been published, and independently re-prove the finding from scratch. The entire effort consumed 31 million output tokens across two sessions in Claude Code.
Perhaps most remarkably, Claude was initially skeptical of its own result. The model had learned from training data about the extreme difficulty of open problems in mathematics and the known limitations of AI systems. It reportedly took encouraging prompts before Claude accepted that its finding was valid. As Anthropic noted in its announcement: "Perhaps Claude, like many of us, underestimates the rate of AI progress."
Human Validation and Formal Verification
The finding did not rest on Claude's word alone. Two mathematicians at Anthropic — Levent Alpöge and Ralph Furman — studied Claude's work and produced an informal technical note for experts. Alpöge had previously made headlines for using Claude to disprove the 87-year-old Jacobian conjecture, a separate mathematical breakthrough reported earlier in August.
Claude also produced a machine-checked formal proof of its result using Lean, a proof assistant that verifies mathematical arguments with software. The formalization, created with Anthropic staff member Eric Easley, passes the standard validation tool called a comparator.
External experts Brian Conrey and Dan Goldston, both recognized authorities in the study of the Riemann zeta function, examined the paper on short notice and validated the approach.
What the Result Means
The Riemann zeta function describes the distribution of prime numbers: each zero of the function contributes successively finer detail to the sequence of primes. The Riemann hypothesis predicts that all the zeros determining the primes lie along a single vertical line in the complex plane. Proving this has become one of the most consequential conjectures in mathematics, with hundreds of results depending on it as an assumption.
While no one has proven or disproven the full hypothesis, mathematicians have made incremental progress by quantifying the minimum proportion of zeros that do lie on the line. Over decades of research, that known constant had been pushed to 41.6%. Claude's improvement to 67.2% is the largest single jump in this measure in years.
Anthropic cautioned that the techniques Claude used are unlikely to lead to a full proof of the Riemann hypothesis. However, the result demonstrates that AI models can extend the reach of mathematicians' ideas in new and sometimes surprising directions.
A Pattern of AI Mathematical Discoveries
The Riemann zeta advance follows a growing list of AI-assisted mathematical breakthroughs. Earlier in 2026, an OpenAI model called Astra solved ten long-unsolved mathematical problems with machine-checked proofs. Anthropic's own Claude Fable 5 had disproved the Jacobian conjecture weeks earlier. These results suggest that frontier AI models are moving beyond pattern recognition and text generation into the territory of genuine mathematical research.
Claude's Riemann zeta finding stands apart in one respect: it was not the product of a directed research program. It emerged when a non-mathematician asked an AI model an unreasonable question, and the model answered with something no human had found before.
Stay Ahead of AI
For the latest breaking AI news, research breakthroughs, and industry analysis, visit AI Buzz Wire.
Read more AI news
