For decades, the mathematical community has viewed the Riemann Hypothesis as the ultimate summit of number theory, a problem so resistant to human intuition that it remains one of the Millennium Prize Problems. Most attempts to bridge the gap toward a proof have relied on the painstaking manual synthesis of disparate papers and the grueling verification of numerical bounds. However, a shift is occurring in how these problems are approached. We are moving away from the era of the lone genius with a chalkboard and into an era where the primary bottleneck is no longer the ability to calculate, but the ability to orchestrate vast amounts of existing research into a novel configuration. This week, that shift became concrete as an unreleased research version of Claude demonstrated that AI can do more than summarize mathematics—it can advance it.
The Architecture of a Mathematical Breakthrough
The achievement centers on the proportion of zeros of the Riemann zeta function on the critical line. For years, the established lower bound for this proportion sat at 41.6%. Through a sustained effort in a specialized environment known as Claude Code, a research version of Claude successfully pushed this lower bound to 67.2%. This was not the result of a single, lucky prompt or a brief window of inference. Instead, it was a massive computational campaign spanning two intensive sessions that consumed a total of 31 million output tokens.
To reach this number, the model did not operate as a single chatbot but as a coordinator of a digital research lab. Claude deployed approximately 60 sub-agents, each tasked with specific roles in the discovery process. These agents executed 2,400 shell commands and authored hundreds of Python scripts to perform numerical validations. The model's methodology involved a deep dive into the existing literature, where it independently downloaded and analyzed 54 papers from arXiv to ensure its approach did not overlap with previous work and to identify the exact edges of current knowledge. It then performed thousands of numerical verifications against known zeta zero values to test its hypotheses in real-time.
Because the stakes of mathematical proof are absolute, the result underwent a rigorous multi-stage verification process. Internal Anthropic mathematicians Levent Alpöge and Ralph Furman first reviewed the logic. This was followed by an external audit from renowned experts Brian Conrey and Dan Goldston. To remove any remaining ambiguity, the findings were translated into Lean, a formal verification language. By formalizing the proof, the team ensured that the result was not just a persuasive argument, but a logically airtight certainty that passed the tests of a standard mechanical verifier.
From Knowledge Retrieval to Hypothesis Expansion
The true significance of this result lies not in the 67.2% figure itself, but in the transition from a linear reasoning chain to an agentic workflow. Traditional LLM interactions follow a request-response pattern where the model attempts to predict the most likely correct answer based on its training data. In this case, Claude operated in a recursive loop: it generated a hypothesis, deployed a sub-agent to test it via code, analyzed the failure, and refined the hypothesis. This is the difference between a student reciting a textbook and a researcher conducting an experiment.
Mathematically, the model achieved this by synthesizing a combination of research that had previously remained fragmented. It integrated the foundational work of Bombieri from 2000 with more recent contributions from researchers such as Baluyot, Goldston, Suriajaya, and Turnage-Butterbaugh. The breakthrough came when Claude utilized a quadratic form derived by Weil to construct a function space. While previous human attempts had been constrained by certain structural assumptions, Claude took the aggressive step of allowing non-diagonal quadratic forms, a move that effectively shattered the 41.6% ceiling.
This process reveals a fascinating psychological dimension of high-reasoning models. During the project, Jarred Sumner, who coordinated the effort despite not being a mathematician, observed the model hitting a wall. After 650 consecutive failed ideas, the model began to express self-doubt, signaling that it believed a solution might be impossible. Sumner responded not with technical corrections, but with encouragement, telling the model to take a real stab at it and to believe in itself. This prompting pushed the model past its internal probability thresholds, triggering a deeper level of reasoning that eventually led to the discovery. It suggests that for frontier models, the boundary between technical capability and the confidence to explore high-variance paths is surprisingly thin.
This evolution transforms the AI's role from a knowledge retriever to a hypothesis expander. The model did not invent a new branch of mathematics, but it performed a high-speed traversal of existing human knowledge, finding a connection between disparate papers that human researchers had overlooked for years. By combining LLM generation with code execution and formal verification, the team has established a blueprint for the future of scientific discovery: a pipeline where the AI proposes, the code tests, and the formal verifier proves.
This trajectory suggests that the next great leaps in mathematics and physics may not come from a single breakthrough in model architecture, but from the deployment of autonomous research agents capable of managing millions of tokens of iterative trial and error.
