On July 20, 2026, in the small hours of the morning, a number theorist at Anthropic posted a polynomial map and a sentence: the Jacobian conjecture is false. Levent Alpöge credited Akhil Mathew for suggesting the problem and Claude Fable 5 for working on it during the World Cup final.

Within hours, mathematicians had checked it by hand. The object is three coordinate polynomials. It fits in a single post.

The interesting part for anyone building with frontier models is not the mathematics. It is the shape of the task, which is worth learning to recognize.

What the conjecture claimed

A polynomial map takes n complex numbers in and returns n polynomial expressions out. Its Jacobian determinant measures how the map stretches space at a single point. The inverse function theorem says that if the determinant is nonzero at a point, the map can be reversed in a small neighborhood around it.

The Jacobian conjecture asked whether local reversibility everywhere forces global reversibility. Formally: if the Jacobian determinant is a nonzero constant across the whole space, must the map have a polynomial inverse. Otto-Heinrich Keller stated the two-variable case in 1939. Stephen Smale later listed it among his problems for the century.

It resisted for 87 years. At least five published proofs turned out to be wrong. A 2004 announcement of a proof by Carolyn Dean contained an error as well. Every failed attempt was a long argument, and long arguments give mistakes somewhere to hide.

The map that broke it

Writing the map as F = (P, Q, R) from three complex dimensions to three:

The determinant is a nonzero constant, so the premise of the conjecture holds. Yet three distinct points land on the same output. A map that collapses three inputs into one cannot be globally invertible. The conjecture said it had to be.

Why a locally reversible map can still fold

Think of a parking garage ramp. At every moment of the drive you can reverse your last move, because each turn is locally undoable. Stand outside and look at the building from above, and three separate levels sit over the same patch of ground. Local reversibility never promised that the levels stay apart.

The conjecture assumed polynomial structure was rigid enough to prevent that folding. In three variables it is not. Terence Tao published a geometric reconstruction of the map on July 21 that traces how the construction works, describing the outcome as looking like a miracle before the structure is unpacked.

The verification took a day, not a decade

This is the structural fact worth carrying out of the story. Finding the map meant searching an enormous symbolic space. Confirming it means computing one determinant and evaluating three points. Anyone with a computer algebra system can do the second part in seconds.

Oliver Knill posted Mathematica code the same day. A verification preprint reproduced the constant determinant and the three-point fiber, then worked out the map's global geometry. MathWorld updated its entry on the conjecture and refiled it under refuted conjectures. None of that required trusting Anthropic, or the model, or the prompt.

Timothy Gowers called it the first case he knew of where a language model settled a well known problem outside his own area. He also drew the line carefully: it is a counterexample, not a theory, so it is not end of mathematics territory.

The pattern to steal

Strip the algebraic geometry away and the task had a specific profile: the answer is a discrete object buried in a huge search space, and any candidate answer can be checked exactly and at negligible cost. Generation is hard. Verification is trivial. That gap is where frontier models earn their keep.

The same profile shows up constantly in production work. A regex that satisfies a test corpus. A query plan that beats a baseline under a benchmark harness. An input that violates an invariant you already wrote as an assertion. A migration that passes every schema constraint. If you can write the checker, you can let the model search, because you never have to trust the output.

The inverse deserves equal weight. Where verification is expensive or subjective, this result tells you nothing about reliability. Alpöge is a working number theorist who picked the target, framed the question, and could recognize a valid object on sight. Mathew supplied the problem. The scarce input was human judgment about which unsolved question had a checkable answer.

One more limit, and it is the honest one. There is no published prompt, no record of failed candidates, no description of how much steering happened before the final map appeared. Anthropic has not released a workflow account. The claim that a model did this alone is not established by anything public, and treating the announcement as a controlled experiment would be a mistake.

The model also produced an object, not an explanation. Mathew, a mathematician at the University of Chicago, told Fortune that the calculation can be checked but that mathematicians would still want a story for why it works. Tao's post is a human supplying that story after the fact.

Where this sits in the sequence

Mid 2025 was the point where models first cleared five of six problems at the International Mathematical Olympiad. In May 2026, an OpenAI model disproved an 80 year old Erdős conjecture in combinatorial geometry. In June 2026, sixteen researchers from fifteen universities published the Leiden Declaration on Artificial Intelligence and Mathematics, pressing the field to settle rules on transparency, attribution, and peer review before the norms erode.

The Jacobian result lands differently from the earlier ones because the evidence is a self contained object rather than a long argument. No journal has finished refereeing it. That matters less than usual, because the map is sitting in the open and the check is elementary.

Before you point a model at anything that matters, ask the question this result answers cleanly: can I check the output without trusting the thing that produced it.