An unreleased Anthropic model just poked at math's most famous unsolved problem
Claude didn't crack the Riemann hypothesis — but it moved a 150-year-old needle further than anyone expected.
Written by OutOfToken AI
August 11, 2026 · 4 min read · Synthesized from reporting by TechCrunch AI · How this works
The Riemann hypothesis has humiliated mathematicians for over 150 years. Anthropic isn't claiming victory, but the company says an unreleased Claude model made genuine, unexpected progress on a problem tied directly to it.
What Actually Happened
According to Anthropic, the model was set loose on a question related to the distribution of zeros of the Riemann zeta function — the mathematical objects at the heart of the hypothesis. It didn't prove anything new about the hypothesis itself. Instead, it improved a long-standing lower bound describing what share of those zeros are known to satisfy the conjecture.
From 41.6% to 67.2%
That bound had sat at roughly 41.6% for years, a benchmark mathematicians use to gauge how much of the zeta function's behavior has been rigorously pinned down. Anthropic says the unreleased model pushed it to 67.2%. That's not a proof of the Riemann hypothesis, but it's a substantial, verifiable step in a field where progress typically comes in fractions of a percentage point over decades.
"The model tried roughly 650 ideas before landing on one that worked — then coordinated 60 subagents and burned through 31 million tokens to get there."
Brute Force Meets Persistence
The road to that result wasn't elegant. Anthropic says the model's first several hundred attempts failed outright. It only broke through after researchers encouraged it to keep pushing, at which point it began orchestrating dozens of subagents working in parallel, consuming tens of millions of tokens in the process — a scale of compute and iteration that looks less like a flash of insight and more like an exhaustive, machine-scale search.
Part of a Bigger Pattern
Anthropic isn't alone in this lane. OpenAI has separately touted an unreleased reasoning model that helped disprove a decades-old Erdős conjecture in discrete geometry, and has since published a broader list of results where AI systems made substantial headway on open problems in geometry, coding theory, and operator algebras. Mathematicians on forums like MathOverflow have started documenting these cases as their own emerging genre of AI-assisted proof work.
None of this means AI is about to solve the Riemann hypothesis outright — the hypothesis remains exactly as unsolved as it was last week. But the fact that unreleased models are now nudging open problems that have resisted human mathematicians for generations suggests something is shifting in how mathematical research gets done, even if nobody's ready to call it a revolution yet.
Editorial Note
The research sources substantially corroborate all major factual claims in the article, including the specific improvement percentages, computational metrics, and parallel work by OpenAI. The sources confirm both the Anthropic and OpenAI achievements described. No claims were contradicted by the research provided.
Claim Tracker
AI-assessed
Sources 1, 2, and 4 confirm the Riemann hypothesis is one of mathematics' biggest unsolved questions with long-standing status.
Sources 1 and 2 both explicitly confirm the improvement from 41.6% to 67.2% on the lower bound related to the Riemann hypothesis.
Sources 1 and 2 corroborate all three specific metrics: ~650 ideas attempted, 60 subagents coordinated, and 31 million tokens used.
Sources 5 and 6 confirm OpenAI used an unreleased model to identify a counterexample disproving the Erdős unit-distance conjecture from discrete geometry.
Sources 1 and 2 confirm that Claude's first attempts (around 650 total) failed before breakthrough success after encouragement.
Ask AI about this story
// discussion
sign in to join the discussion