Claude Solved a 60-Year Math Problem. What Changes Next

A 60-Year Puzzle, Solved by a System That Wasn’t Trained for It

Claude proved the dying percolation conjecture this week. Not a hint toward it, not a numerical approximation—it delivered a complete formal proof in Lean, open-sourced on GitHub under the repository anthropics/formal-math. The conjecture had resisted some of the best probabilists alive for six decades.

The immediate reading is that AI can now solve problems that defeated human experts. The more interesting one is about what that implies for how research actually gets done.

Why This Dimension Range Was Impossible

Percolation theory dates to 1957, when Simon Broadbent and John Hammersley modeled fluid flow through porous materials. The central object is simple: take a lattice Z^d, open each edge independently with probability p, and ask whether an infinite connected cluster emerges. There is a critical threshold p_c below which all clusters stay finite and above which one diverges infinitely. The question is what happens exactly at p_c.

The “dying” percolation conjecture asks whether the probability of an infinite cluster at that threshold is always zero, regardless of dimension. In two dimensions, Harry Kesten proved in 1980 that p_c equals one-half and the conjecture holds. Above ten dimensions, lace expansion techniques make the problem tractable—space is so vast that two clusters almost never intersect by accident. The stubborn gap was everything in between: dimensions three through ten.

These resist both available tools. Two-dimensional geometry provides no foothold in higher lattices. Ten-dimensional independence breaks down. The intermediate range is simultaneously too structured for probabilistic approximations and too chaotic for geometric intuition. It remained the hardest regime precisely because it was the one most relevant to physical reality.

What Claude Actually Proved

Claude did not attack the full conjecture from scratch. The proof rests on a 2024 formulation by Gady Kozma and Shahaf Nitzan, who proposed a family of gluing inequalities valid on arbitrary finite weighted graphs. Proving any member of that family would imply theta(p_c) equals zero across all dimensions d greater than or equal to two.

Kozma and Nitzan had already shown that even the weakest member—the so-called near-one gluing inequality—was sufficient. Claude proved that weak version.

The 15-page Lean formalization includes a structural guide and sits in the same repository containing Claude’s 11-day autonomous formalization of Fermat’s Last Theorem. Neither the near-one gluing proof nor the Fermat work was presented as a solo feat. Both build explicitly on prior human mathematics. The difference is that the decisive step—a step no human had managed in 60 years—came from the model.

The Field Has Been Expecting This

Hugo Duminil-Copin, winner of the 2022 Fields Medal for work on phase transitions in percolation, published an essay titled “Care for a little more AI?” on his blog Proofs and Prompts on August 30, 2026, just days before the announcement. He identified the dying percolation conjecture as the most famous open problem in the area and wrote that an artificial intelligence would reach a solution before a human would.

Duminil-Copin was not surrendering. His argument was that model-generated proofs represent only the beginning of mathematical digestion. Someone still has to reconstruct the conceptual path, situate the result in its history, generalize it, and translate it into language other humans can absorb and extend. His own failed attempts at the conjecture, he noted, had already yielded independent insights.

The question he posed is the one the community is now living: as research flows increasingly through models, what happens to the accidental discoveries that come from human struggle rather than systematic search?

Reactions Split Along a Clear Fault Line

Benedikt Jahnel at Technische Universität Braunschweig, who called the conjecture the field’s “holy grail” and observed that a human solution would plausibly merit a Fields Medal, described mixed emotions after the announcement. He expressed genuine delight at the result alongside a deflation that the decisive advance came from a system rather than a person.

The news spread quickly through the community. Itai Benjamini flagged it to Gil Kalai, who posted about it on his blog. Jeff Steif shared a more personal reflection on what a lifetime of percolation work means when the hardest step is taken by something else. Alonso Castillo-Ramirez argued that the beauty of a proof does not depend on who discovers it.

The underlying question is identical across all responses. When someone—or something—crosses a threshold you spent decades approaching, what remains for you to do?

This Is Not the First Time

Percolation is the latest entry on a short list. In August, Anthropic announced that Claude had improved bounds related to the Riemann hypothesis. Before that, a model in the same family solved ten previously unsolved problems with Lean-verified proofs. The repository also contains the Fermat formalization.

The pattern has not been uniformly clean. OpenAI claimed in a separate case to have disproven an Erdős conjecture on a distance problem after 80 years, then faced criticism for unattributed borrowing from other researchers. It also announced a Navier-Stokes result that drew scrutiny. In September 2026, 25 Fields Medal recipients issued a joint warning about AI use in mathematics. The resulting debate escalated until Terence Tao’s cautious position drew public criticism from Salvatore Sanfilippo, known online as antirez.

The percolation proof occupies a different place in this landscape. It was published, formally verified in Lean, and built explicitly on known conjectures. Controversy over its correctness should be minimal. The controversy is entirely forward-looking.

What Changes Now

The proof does not replace the human mathematician. It replaces the human who would have been first to the critical step.

That distinction matters. Mathematical research has always involved a relay: earlier work defines the terrain, identifies the obstacles, and passes the baton forward. Claude ran the leg that no one else could. The question for the field is whether receiving the baton faster changes the value of the run itself.

Jahnel’s concern captures the tension. The result is correct. The proof checks out in Lean. But the process that produced it—a system optimizing toward a formal verification target without the messy detours, false starts, and side discoveries that come from sustained human engagement—may omit something irrecoverable. The auxiliary ideas that emerge from failure, the reorientation of intuition, the contextual understanding that makes a proof meaningful rather than merely valid. Those are what Duminil-Copin was gesturing toward when he described digestion as necessary work, not optional decoration.

The near-one gluing inequality was the key. It was discovered and proposed by humans. Claude found the path through it. That sequence—human insight defining the structure, machine execution closing the gap—may be the durable model rather than an anomaly.

What comes next depends on whether the community treats this as a finish line or a method. The conjecture is resolved. The conversation about how research operates inside and alongside these systems is just beginning.