Unreleased Claude model advances math proof

An unreleased Anthropic model has pushed the lower bound for where any counterexample to the Riemann hypothesis would have to lie, Anthropic said Monday. The hypothesis, unsolved since 1859, carries a $1 million Clay Mathematics Institute bounty that remains unclaimed. Anthropic is not claiming a proof: the model narrowed where a violation could theoretically exist, not whether one does.

Getting there took roughly a day and a half of work: the model tested about 650 distinct ideas across 60 coordinated subagents, burning through 31 million output tokens. Two of Anthropic’s in-house mathematicians checked the result, and the team formalized it using Lean, the open-source proof assistant that turns a claimed proof step into something a computer can verify line by line rather than take on faith.

What it means for you

The headline isn’t that a chatbot solved a Millennium Prize problem, it didn’t. It’s that a research-stage model can now run a multi-day, multi-agent search over a genuinely hard open problem and produce output rigorous enough to survive formal verification, not just a plausible-sounding writeup. If your organization is exploring running coding agents remotely or getting results from multi-hour agents on harder analytical work, this is a useful data point on how far that pattern currently scales, provided you still pair it with the kind of independent, formal verification Anthropic applied here rather than trusting the output on its face.