Let's get one thing straight: the Riemann Hypothesis, arguably the most famous unsolved problem in mathematics, is still unsolved as of September 2026. No AI has claimed the one million dollar prize offered by the Clay Mathematics Institute. But what happened in August 2026 might be more significant than if an AI had simply spat out a proof. An unreleased research version of Anthropic's Claude model made unexpected progress on a related mathematical bound. A non-mathematician at Anthropic, Jarred Sumner, prompted the model with an audacious request: take a serious shot at proving the Riemann Hypothesis. The AI couldn't do it. Nobody expected it to. But while exploring the problem over a day and a half, Claude improved a longstanding lower bound that had stood as the best human result for decades. The bound in question measures what fraction of the zeros of the Riemann zeta function are guaranteed to lie on the critical line where the hypothesis predicts all of them should be. Mathematicians had proven that at least 41.6% of these zeros satisfy the hypothesis. Claude pushed that number to 67.2%. To get there, the AI tested 650 different approaches, coordinated roughly 60 sub-agents, performed thousands of numerical checks, wrote hundreds of Python scripts, and searched through 54 research papers. James Maynard, a mathematician at the University of Oxford, called the result genuinely impressive, saying the problem needed a new real idea, which Claude's result appears to provide. Here's what makes this wild: the AI was initially skeptical of its own finding, possibly because it had learned from training data about the difficulty of open mathematical problems and the limitations of AI models. After some encouraging prompts from humans, it arrived at the verified result. Two mathematicians at Anthropic validated Claude's paper, and external experts Brian Conrey and Dan Goldston examined it on short notice. Claude also produced a formally verifiable proof using Lean, a proof assistant system that checks mathematical arguments line by line. This wasn't an isolated incident. August 2026 marked a inflection point for AI in mathematics. On August 1, 2026, OpenAI announced that Astra, its internal model, had solved 10 longstanding open problems in mathematics and theoretical computer science. In July 2026, Levent Alpöge, a mathematician at Anthropic, disproved the Jacobian conjecture using Claude Fable 5. Multiple Erdős problems have fallen to AI models throughout 2026. Google DeepMind's Gemini Deep Think scored gold medal standard at the 2025 International Mathematical Olympiad and released an AI co-mathematician system in May 2026 that scored 48% on FrontierMath Tier 4, problems designed to stump AI for decades. The workflow that produced Claude's Riemann result reveals the future of mathematical research. This wasn't a genius flash of insight. It was brute-force exploration at scale, kept alive by human encouragement. Out of Claude's 60 sub-agents, only two developed the key mathematical ideas. Thirteen contributed supporting ideas. Thirty tried and failed. The process burned through 31 million output tokens. What emerged wasn't elegant mathematics in the classical sense but rather a patient, massive computational search that ground through hundreds of dead ends before finding something that worked. Anthropic is careful to note that the techniques Claude used are unlikely to lead to proving the Riemann Hypothesis itself. The hypothesis, proposed by Bernhard Riemann in 1859, concerns the distribution of prime numbers along the number line. It predicts that all non-trivial zeros of the Riemann zeta function have a real part equal to one-half. Proving it would revolutionize number theory and have applications in cryptography and countless other fields. But a full proof, experts agree, will require a fundamentally new approach, not incremental improvements on existing bounds.