Anthropic mathematicians validated the result, and outside experts in number theory reviewed the findings, which include a formally verifiable proof.
This breakthrough suggests that AI models are becoming capable of synthesizing complex, existing research to produce original mathematical contributions.
To reach the new bound, the model coordinated approximately 60 subagents—specialized instances of the AI working together—to write hundreds of scripts and execute thousands of numerical checks.
This process allowed the AI to combine decades of human-led research in ways that previously had not been successfully connected, demonstrating a significant leap in the computational and reasoning speed of AI infrastructure.
The discovery was an unintended byproduct of a staff member's request for the AI to attempt the full Riemann hypothesis.
Over the course of two sessions, the model used 31 million output tokens and verified its own work by cross-referencing dozens of academic papers to ensure the finding was original.
While the specific techniques used are not expected to solve the centuries-old Riemann problem entirely, the achievement marks a shift in how AI can be used to extend the reach of human academic research in highly technical fields.