Anthropic says an unreleased model has made notable progress on the Riemann hypothesis, one of mathematics' most enduring open questions. The problem, which has challenged researchers for more than 150 years, is closely tied to the pattern of prime numbers and remains one of the field's most famous unsolved puzzles.
According to the company, the model improved the known lower bound for cases where the hypothesis holds. What makes the result especially interesting is the way it was achieved: a staff member with limited mathematical background asked the system to take a serious attempt at the proof, then let it continue working across roughly a day and a half.
During that period, the system explored 650 ideas, coordinated 60 sub-agents, and used 31 million tokens in total. Anthropic's report says only a small group of those sub-agents generated the key mathematical insights, while others tested, validated, and helped shape the final write-up. The result was then checked by in-house mathematicians and formalized with the Lean proof assistant.
The development adds to a growing wave of AI-assisted mathematics. Recent months have seen large language models contribute to several advanced problems, including results from OpenAI's internal research and other work from Anthropic. For many researchers, these examples suggest that AI is becoming a serious partner in mathematical discovery.
At the same time, the progress is prompting broader discussion about authorship, verification, and how future proofs may be credited. As AI systems become more capable of exploring abstract problems, they may help reshape how mathematics is researched, validated, and shared in the years ahead.
