On July 20, Anthropic researcher Levent Alpoge announced a remarkable achievement: using the Claude Fable 5 language model, he managed to find a counterexample to the Jacobian conjecture. This problem, formulated by German mathematician Ott-Heinrich Keller in 1939, is one of the central unsolved problems in algebraic geometry, which the world's best minds have grappled with for 87 years.
The Jacobian conjecture, simply put, states that if a locally polynomial mapping never loses information (its Jacobian never vanishes), then it is globally invertible. Claude Fable 5 constructed a specific formula that passes the local test at every point, yet maps three different sets of numbers to the same result. This makes it impossible to recover the original data, thus proving the conjecture false. Alpoge provided links to Wolfram Alpha to verify the calculations.
The context of the discovery is particularly interesting: according to the researcher, the question about the conjecture was asked by a friend, and the model itself worked on the problem during the World Cup final. The Jacobian conjecture is notoriously known for attracting erroneous proofs. Mathematician Qiaochu Yuan noted that this is the most famous open problem solved by a language model and cited T. T. Moh, who in 2008 analyzed one of the false proofs, mentioning that even such authorities as Beniamino Segre and Claude Chevalley made mistakes in their attempts.
Mathematician Jared Duker Lichtman called the result "outstanding," recalling that a special case of the conjecture was the subject of Ethan Zhang's dissertation — the future author of a breakthrough in prime number theory. It is important to emphasize that the result has not yet undergone peer review, but its main strength lies in verifiability. Unlike AI companies' claims about benchmark records, a counterexample is a formula that any mathematician can independently plug in and recalculate, making the discovery transparent and verifiable.
Alpoge earned his PhD from Princeton under the supervision of Fields Medalist Manjul Bhargava and received the Morgan Prize in 2015 — the highest US award for undergraduate research in mathematics. His work with Claude demonstrates that AI is capable not only of solving standard problems but also of finding elegant, non-trivial refutations that have eluded humans for decades.
AI vs. Open Problems: A New Era of Mathematics?
This is not the first time language models have solved mathematical problems. In October 2025, OpenAI Vice President Kevin Weil claimed that GPT-5 had solved ten Erdős problems, but the claim was not confirmed — the model found existing solutions. Later, in January 2026, GPT-5.2 Pro solved Erdős problem No. 397, which was confirmed by Fields Medalist Terence Tao. In May, an OpenAI model disproved Erdős's 1946 conjecture on unit distances, and Google DeepMind introduced AlphaProof Nexus, which closed nine Erdős problems.
The Star Fleet Math project in July presented solutions to 20 Erdős problems obtained by 20 parallel Codex agents. In February, Google launched the AI mathematician Aletheia, which autonomously solved four problems from the Erdős list. However, the counterexample to the Jacobian conjecture stands apart: it is not just solving another problem, but a refutation of a fundamental conjecture that was a pillar of algebraic geometry.
Expert opinion: This event is not just a victory for AI in mathematics, but a demonstration of a paradigm shift. Previously, we saw models solving complex but isolated problems. Here, we observe AI acting as a full-fledged researcher, capable of finding weaknesses in long-standing theoretical constructs. For the crypto industry, this is a signal: the mathematical rigor and verification underlying blockchain could gain a powerful tool for automatically finding vulnerabilities and proofs. If AI can refute 80-year-old conjectures, then checking smart contracts or cryptographic protocols could become a matter of minutes, not weeks of auditing.