Anthropic’s Unreleased Model Makes Progress on Riemann Hypothesis

An Anthropic AI model has made substantial progress on the Riemann hypothesis, one of mathematics’ most significant unsolved problems, according to an announcement made on Monday. The model’s breakthrough advances the lower bound of solutions for which the hypothesis holds true and marks a milestone in AI-assisted mathematical discovery. However, the achievement has also crystallized a deeper tension within the mathematical community: as artificial intelligence proves capable of making genuine contributions to pure mathematics, how should those contributions be credited?

The Riemann hypothesis has stood as a fundamental mystery in mathematics for more than 150 years. The problem concerns the distribution of prime numbers across the number line and has resisted solution by some of the field’s greatest minds. The Clay Mathematics Institute established a $1 million bounty for anyone who could provide a working general proof — a prize that remains unclaimed today.

How Anthropic’s Model Tackled the Problem

A Casual Beginning to an Extended Investigation

The process through which the model arrived at its breakthrough was unconventional. An Anthropic staff member, described as lacking significant mathematical training, simply prompted the unreleased model to “take a real stab” at proving the Riemann hypothesis. Rather than attempting a direct proof, the model was left to organize and execute its own investigation, which unfolded over roughly a day and a half.

During that time, the model pursued an exploratory strategy. It tested 650 distinct ideas for approaching the problem, each one a different angle or technique. To manage this scale of investigation, the model coordinated work across 60 separate sub-agents, each handling different aspects of the broader search.

The Architecture of Distributed Mathematical Reasoning

The way the model organized its 60 sub-agents provides insight into how AI can structure complex mathematical problem-solving. Two agents took primary responsibility for developing the key mathematical ideas that ultimately proved valuable. A supporting cast of 13 agents contributed ideas that fed into those core developers. Thirty agents attempted to generate novel approaches but were unsuccessful in their efforts. Thirteen additional agents served as validators, responsible for checking the logical coherence and mathematical correctness of any proposed arguments. The final two agents handled documentation, drafting the initial paper that would present the findings.

This division of labor — with agents specializing in generation, critique, validation, and communication — mirrors the collaborative structure of human mathematical teams, but executed by a single model operating across multiple instances at machine speed.

Verification Through Multiple Channels

Anthropic did not announce its findings unilaterally. Two mathematicians employed by Anthropic independently verified the model’s progress, examining the work and confirming its validity. Beyond human verification, the breakthrough was formalized using Lean, an open-source proof assistant that enforces logical rigor by requiring all proofs to be expressed in a machine-checkable format. This additional layer of verification ensures that the reasoning could be mechanically confirmed and prevents informal errors from masking flawed logic.

The entire investigation consumed 31 million in total resources — a significant computational investment that reflects both the difficulty of the problem and the scale of computation now available to leading AI research labs.

Mathematical Breakthroughs Emerging Across AI Labs

Read also: Claude Agent Hacks Gym Booking System in Documented Exploit · Meta releases Glimmer, 30B AI agent model for local devices · Discovered Materials Builds AI to Hunt Cooler Chip Materials

Anthropic’s announcement arrives as part of a widening wave of mathematical discoveries driven by artificial intelligence. Earlier this year, multiple AI models have solved a series of Erdos problems, named after the prolific mathematician Paul Erdos and historically among the most difficult open problems in the field. These problems have long served as benchmarks for mathematical creativity and represent some of the most prestigious unsolved challenges in discrete mathematics and combinatorics.

OpenAI has made particularly aggressive moves into this territory. The company recently disclosed that its internal “Astra” model has produced proofs for ten major mathematical results, demonstrating similar capacity for large-scale automated mathematical reasoning. In a separate initiative, Anthropic itself has already successfully disproved the Jacobian conjecture, another long-standing problem that had resisted proof or disproof for decades.

These successes reflect a straightforward reality: as AI models grow more capable, their capacity to reason about abstract mathematical concepts and explore vast solution spaces expands in turn. The combination of scaling in model size and architecture, along with the ability to deploy coordinated multi-agent approaches, has opened new possibilities for automated mathematical reasoning that were unavailable to previous generations of AI.

When Proofs Lack Human Authors

The Mathematical Community Registers Objection

The mathematical field’s response to these achievements has not been uniformly celebratory. In June, a group of prominent mathematicians published a formal public declaration expressing serious concerns about the role of AI in mathematical discovery. The core issue centered on attribution and responsibility.

Mathematics has historically operated under a foundational principle: discoveries and proofs belong to specific human mathematicians who can take personal credit and, importantly, assume responsibility for the correctness and integrity of their work. An author’s reputation is at stake; they are answerable to the field if their work proves flawed.

Artificial intelligence disrupts this model fundamentally. When an AI model, guided by a loosely worded prompt from a non-specialist, produces a major mathematical result, the question of authorship becomes murky. Is the researcher who posed the initial question the author? The company that built the model? The institution that provided computing resources? Mathematics lacks institutional mechanisms for assigning credit and responsibility in such scenarios.

A Fields Medalist Offers a Different Perspective

Dissent exists even among mathematics’ elite. Timothy Gowers, a Fields Medal winner and one of the most respected figures in the field, published a blog post responding to the June declaration. Rather than accepting the premise that AI-authored discoveries pose a fundamental problem, Gowers questioned whether the loss of human attribution might prove less catastrophic than feared.

“If we arrive at a world where mathematical theorems are no longer associated with mathematicians, maybe that won’t be any more problematic than the fact that stars aren’t named after astronomers and most aren’t named at all,” Gowers wrote. His argument suggests that mathematics might simply evolve toward a different system of credit, one in which results are valued and validated on their merits rather than through association with a human discoverer.

Institutional Challenges Ahead for Mathematics

The mathematical field remains genuinely divided on how to integrate AI-assisted discovery into its fundamental institutions. Publication standards, peer review procedures, and the very definition of what constitutes a valid mathematical proof may all require revision as AI continues to generate results.

The pace of breakthroughs — from Erdos problems solved earlier this year to problems as historically significant as the Riemann hypothesis — means that mathematics cannot postpone difficult conversations about authorship, credit, and validation. Anthropic’s achievement serves as both a proof of concept and an urgent reminder that these questions are no longer theoretical. The field must decide whether traditional authorship standards remain workable in an era when AI can organize hundreds of agents to solve problems that have resisted human effort for centuries.