An Unreleased Anthropic Model Made Progress on the Riemann Hypothesis
An unreleased Anthropic model has made measurable progress on the Riemann hypothesis, one of mathematics' longest-standing unsolved problems. The model increased the lower bound of zeros confirmed to lie on the critical line. It did not solve the hypothesis generally.
The $1 million Millennium Prize remains unclaimed. The hypothesis has been open for more than 150 years.
How It Happened
An Anthropic staff member without significant mathematical training prompted the model to "take a real stab" at the Riemann hypothesis. The model then ran autonomously for a day and a half.
It tested 650 distinct ideas. It coordinated 60 subagents and spent 31 million output tokens in total. Of the 60 subagents: 2 developed the key mathematical ideas, 13 contributed supporting ideas to those agents, 30 attempted and failed to develop new ideas, 13 acted as validators, and 2 helped draft the initial paper.
Two Anthropic in-house mathematicians confirmed the result. The proof was then formalized using Lean, the open source proof assistant.
Context: A Busy Year for AI and Mathematics
This is not an isolated result. Multiple Erdos problems have been solved by AI models in 2026. OpenAI released 10 major mathematical results from an internal model it calls Astra. Anthropic separately disproved the longstanding Jacobian conjecture.
The pattern is worth noting: these are not proofs of famous open problems requiring entirely new mathematics. They are incremental advances, disproofs, and resolutions of problems that were hard but tractable given enough compute and coordination.
The Riemann hypothesis result fits the same mold. A partial bound improvement, confirmed and formalized, produced by a model no human mathematician was driving in any technical sense.
What This Means in Practice
The subagent breakdown is the most informative part of this announcement. Thirty of 60 agents failed to contribute useful ideas. Two did the real mathematical work. The rest supported, validated, and wrote. That failure rate is high. It also produced a result that advances a 150-year-old problem.
Anthropic has not released the model. The result has not been peer reviewed through standard journals. Whether this generalizes to harder problems, or whether the current approach runs out of headroom quickly, is not yet clear.
For now: a non-mathematician asked a model to try something that has stumped professional mathematicians for a century and a half. The model made measurable progress. That is the fact. Everything else is still being worked out.
Source: Techcrunch