On July 20, 2026, Anthropic researcher Levent Alpoge announced a truly sensational result. Using the Claude Fable 5 model, he managed to find a counterexample to the Jacobian conjecture — one of the central unsolved problems in algebraic geometry, first formulated in 1939 by German mathematician Ott-Heinrich Keller.

The Essence of the Problem and the Refutation

Simply put, the Jacobian conjecture states: if a certain polynomial function of several variables is locally (at every point) invertible and does not "glue" points together, then it is globally invertible. In other words, if no information is lost locally, the process should be reversible overall. This intuitively appealing idea turned out to be false.

Claude Fable 5 constructed a specific formula that brilliantly passes the local test at every point, yet maps three different sets of numbers to the same result. This makes global recovery of the original data impossible, thereby proving the conjecture false. Alpoge included links to Wolfram Alpha for independent verification of the calculations — anyone can check the correctness of the counterexample.

Historical Context and Community Reaction

Over its 87-year history, the Jacobian conjecture has accumulated numerous erroneous proofs, making it famous not only for its difficulty but also as a "trap" for mathematicians. Mathematician Qiaochu Yuan noted that this is undoubtedly the most famous open problem ever solved by a language model. He cited an analysis of one erroneous work from 2008: Beniamino Segre published three incorrect proofs, Claude Chevalley accepted one as correct, and Igor Shafarevich used the conjecture as a proven theorem.

Mathematician Jared Duker Lichtman called the result "outstanding," emphasizing that a special case of this conjecture was the topic of Ethan Zhang's dissertation — the future author of a breakthrough in prime number theory. The result has not yet undergone formal peer review, but its main strength lies in its verifiability. Unlike claims from AI companies based on closed benchmarks, a counterexample is an open formula accessible for verification by anyone.

AI vs. Open Problems: A New Era?

This is not the first time language models have achieved success in mathematics. In October 2025, OpenAI's vice president claimed that GPT-5 had solved ten of Erdős's problems, though it later turned out that some "solutions" already existed in the literature. However, in January 2026, GPT-5.2 Pro successfully solved Erdős problem No. 397, and in May, an OpenAI model refuted the 80-year-old Erdős conjecture on unit distances — the proof was verified by nine external experts. Google DeepMind also introduced the AlphaProof Nexus system, which closed nine smaller problems.

My expert analysis: This case marks a transition from "toy" problems to solving fundamental, centuries-old mathematical challenges. Importantly, the counterexample to the Jacobian conjecture is not merely "found" but is a verifiable and reproducible result. This increases trust in AI as a tool for mathematical discovery. However, as Terence Tao rightly noted, AI is still only picking "low-hanging fruit." Nevertheless, the speed at which models are moving from solving Erdős problems to refuting century-old conjectures suggests we are on the threshold of a new era in mathematics, where AI will become a full-fledged partner for researchers.