Claude's new mathematics story should not be read as an AI solving a century-old problem in one dramatic jump.
On August 10, Anthropic published research material about an unreleased Claude model that was asked to take a serious attempt at the Riemann hypothesis. It failed to solve the hypothesis, but Anthropic says it improved a related lower-bound result: the proportion of Riemann zeta zeros known to satisfy the hypothesis moved from 41.6% to 67.2%.
Set the boundary first
The boundary matters. The Riemann hypothesis remains open. Claude's result concerns a narrower question: how large a fraction of zeros can be proven to lie on the critical line. For a general reader, the plain version is that AI did not open the main door, but it found a new route through a side door mathematicians have studied for years.
Anthropic's evidence stack has several layers: a research post, a Claude-written technical paper, a concise expert note, and a Lean formalization that passes a standard checker. The company says two Anthropic mathematicians studied and validated the work, and outside number theorists Brian Conrey and Dan Goldston reviewed it on short notice.
The original prompt was simple:
“Take a real stab at the Riemann hypothesis.”
In practice, Claude's route looked less like a single answer and more like a research workflow.
Where 67.2% comes from
The key figures need separate labels: 41.6% is the previous lower bound cited by Anthropic; 67.2% is Claude's new lower bound based on prior work by Baluyot, Goldston, Suriajaya, Turnage-Butterbaugh and Bombieri; 1859 and the million-dollar Clay prize describe the larger unsolved conjecture, not the bound itself.
The engineering figures show how the result was produced. Anthropic says Claude first tried 650 ideas without success. It then spent a day and a half coordinating roughly 60 Claude subagents, running 2,400 shell commands, writing hundreds of Python scripts and downloading 54 arXiv papers to check whether the result already existed. The run used 31 million output tokens.
Mathematicians still matter
That makes the human role more visible, not less. The model expanded the search space and drafted proof material. Mathematicians judged whether the route connected to existing work. Lean turned part of the argument into a machine-checkable object.
The next research-agent test
The next things to watch are concrete: independent public review by mathematicians, how much of the proof is covered by the Lean formalization, and whether this workflow produces fresh results on other problems. If it does, the story is not an AI oracle. It is research labor being reorganized around models, scripts, papers and formal verification.
Sources: Anthropic Research, Claude technical paper, Lean formalization repository, Clay Mathematics Institute, CocoLoop, related arXiv papers; verification covers the 41.6% and 67.2% lower bounds, 31 million output tokens, 650 ideas, 60 subagents, 2,400 shell commands, 54 papers and Lean comparator status.