- Anthropic revealed that an unreleased model made significant progress on the Riemann hypothesis, unsolved for more than 150 years, by raising the lower bound of solutions for which it holds true.
- A staff member with no serious mathematical training simply prompted the model to “take a real stab” at it, then left it running for roughly a day and a half.
- The run tested 650 different ideas across 60 coordinated subagents and burned 31 million output tokens; two in-house mathematicians confirmed the result and it was formalized in the Lean proof assistant.
- It joins a 2026 surge of AI math results — solved Erdos problems, OpenAI’s “Astra” proofs, a disproved Jacobian conjecture — reigniting a fierce debate over credit and authorship.
31 million tokens. 60 subagents. 650 separate attempts. That is what it took for an unreleased Anthropic model to make real headway on a problem that has resisted the world’s sharpest mathematicians since Bernhard Riemann first posed it in 1859. On Monday, Anthropic said the model did not solve the Riemann hypothesis — nobody has — but it pushed the frontier further than most people thought a language model could, and it did so with startlingly little human hand-holding.
What Anthropic Actually Did (and Didn’t) Do
The Riemann hypothesis is one of the deepest open problems in mathematics: a 150-year-old conjecture about how prime numbers are distributed. It is one of the Clay Mathematics Institute’s Millennium Prize Problems, carrying a $1 million bounty for a working general proof that remains unclaimed to this day.
Anthropic’s model did not claim that prize. What it did, according to the company’s announcement, was significantly increase the lower bound of solutions for which the hypothesis is known to hold true — a concrete, checkable improvement rather than a hand-wave. The finding was confirmed by two of Anthropic’s in-house mathematicians and formalized using the open-source proof assistant Lean, which mechanically verifies that each logical step is valid.
Trend Insight — The headline is not “AI solved the Riemann hypothesis.” It’s that a general-purpose model produced a genuine, formally verified increment on a frontier math problem. That shifts the question from “can models do math?” to “how much of the grind can they now absorb?”
How 60 Subagents Cracked the Workflow
One vague prompt, roughly 36 hours of autonomy
The most striking detail is how little steering was involved. An Anthropic staff member without significant mathematical training prompted the model to “take a real stab” at proving the hypothesis, then simply left it to coordinate the task over the following day and a half. In that window the model tested 650 different ideas for cracking the problem, orchestrating the work across 60 subagents and spending 31 million output tokens in total.
A division of labor among machines
A footnote to the paper breaks down exactly how those 60 subagents split the work: two were responsible for developing the key mathematical ideas, 13 contributed ideas to those two, 30 attempted but were unable to develop new ideas, 13 served as validators checking the correctness of the arguments, and the final two helped write the initial paper. In other words, the model didn’t just “think harder” — it built itself a research team, complete with idea generators, reviewers, and writers.
Trend Insight — This is a preview of where agentic AI is heading: value comes less from a single brilliant answer and more from orchestration — spawning specialized subagents, discarding 30 dead ends, and keeping 13 validators honest. That structure looks a lot like how human research groups already operate.
Part of a Bigger 2026 Math Wave
The Riemann result is not an isolated fluke. It is the latest in a string of mathematical breakthroughs led by large language models this year. A number of Erdos problems have been solved by AI models over the course of 2026, and each more powerful model release has produced more impressive results.
OpenAI recently published a set of 10 major results proved by its internal “Astra” model, while a separate Anthropic effort disproved the longstanding Jacobian conjecture. Taken together, these are no longer party tricks on competition problems — they are contributions to open, research-grade mathematics, arriving at a pace that even optimists did not expect a year ago.
Trend Insight — When multiple frontier labs independently start clearing research-level math in the same year, it stops being a stunt and starts being a capability curve. Expect “AI-assisted proof” to become a normal line in mathematics papers surprisingly fast.
The Fight Over Who Gets the Credit
The growing body of results has caused both excitement and unease in the mathematical community. In a public declaration signed in June, a group of prominent mathematicians — the Leiden Declaration — warned that AI could undermine core values of the field, particularly the standard that true mathematical proofs should be “attributable to specific authors who take credit for their discovery and assume responsibility for their correctness.”
Not everyone is alarmed. Responding to the declaration in a July blog post, Fields Medal winner Timothy Gowers questioned whether AI might change mathematics in a more complex and even positive way. “If we arrive at a world where mathematical theorems are no longer associated with mathematicians,” Gowers wrote, “maybe that won’t be any more problematic than the fact that stars aren’t named after astronomers and most aren’t named at all.”
Trend Insight — The technical milestone and the cultural fight are arriving simultaneously. The unresolved question isn’t whether AI can help prove theorems — it clearly can — but who is accountable when a machine’s argument is wrong, and what “authorship” even means once a model does most of the reasoning.
Related
- Meta Just Made the Cloud Optional for AI Agents
- Claude Code Will Stop Asking Your Permission Soon
- AI가 만든 화면, 왜 티 날까 (Tech Digest)
- AI 모델 만들다간 망합니다 — 추론으로 옮겨가는 판
Sources
- TechCrunch — An unreleased Anthropic model made progress on one of math’s biggest unsolved problems
- Anthropic Research — Progress on the Riemann hypothesis
- Clay Mathematics Institute — The Riemann Hypothesis (Millennium Prize)
AI Biz Insider · AI Trends EN · aibizinsider.com
댓글 남기기