Claude Fable 5 Killed an 87-Year-Old Math Conjecture

Key Takeaways

– Claude Fable 5 helped a mathematician find a counterexample to the Jacobian conjecture, open since 1939 and ranked among the hardest math problems for the 21st century.
– The counterexample is a degree-7 polynomial map in three variables with constant Jacobian determinant -2 that sends three distinct inputs to the same output. The suspected minimum degree was around 100.
– The conjecture is now false for all dimensions n >= 3. The two-dimensional case remains open.
– The result is hand-checkable with free algebra software and not yet peer-reviewed as of July 2026.

An 87-year-old problem in algebraic geometry fell last weekend. And the tool that helped do it was Claude Fable 5. A mathematician posted a counterexample to the Jacobian conjecture on July 19, 2026. He credited Claude Fable 5 as his collaborator. The math community did what it always does with claimed proofs of famous conjectures: they checked it themselves.

Except this time, checking took minutes, not months. Because the counterexample is a concrete formula anyone can verify with free software.

If you build AI products for a living, the verifiability of this result is a bigger story than the conjecture itself.

That tells you something about where we are with AI capability claims in 2026.

What Was the Jacobian Conjecture?

The Jacobian conjecture asks a deceptively simple question about polynomial maps, which are functions built from sums and products of variables.

If you have a polynomial map from N-dimensional space to N-dimensional space. And its Jacobian determinant (a measure of how the map locally distorts volume) is a nonzero constant everywhere, must the map have a polynomial inverse?

Intuition says yes. A constant nonzero Jacobian means the map never collapses space locally, so it should be globally invertible. Ott-Heinrich Keller formalized this guess in 1939 for N variables, building on earlier two-variable versions of the problem. The conjecture resisted nearly every algebraic geometer who touched it over the following 87 years. It landed at Problem 16 on Stephen Smale’s list of 18 problems for the 21st century, alongside the Riemann Hypothesis and P versus NP.

For nearly nine decades, the answer was “probably yes, but nobody can prove it.”

The answer turned out to be no.

How Did Claude Fable 5 Find the Counterexample?

According to multiple accounts of the discovery, the mathematician was watching the World Cup final on July 19 when a friend asked about the Jacobian conjecture. He opened Claude Fable 5 and started searching for a counterexample. The model proposed candidate polynomial maps. He verified them with symbolic algebra tools.

One of them worked.

The counterexample is a polynomial map F from three-dimensional complex space to three-dimensional complex space. Its Jacobian determinant is the constant -2 at every point. That is a nonzero constant, which is exactly the condition the conjecture requires.

Yet the map is not invertible.

Here is the proof in three lines.

Three distinct input points all map to the exact same output:

– (0, 0, -1/4) produces (-1/4, 0, 0)
– (1, -3/2, 13/2) produces (-1/4, 0, 0)
– (-1, 3/2, 13/2) produces (-1/4, 0, 0)

An invertible function cannot send multiple inputs to the same output. So F cannot have a polynomial inverse.

The conjecture fails for three dimensions and, by extension, for every dimension n >= 3.

The polynomial has total degree 7.

Researchers had previously suspected that any counterexample would need degree around 100. Degree 7 is not supposed to kill a Smale problem. But the math does not care about your expectations.

Why Can You Verify This With Free Software?

Most AI capability claims are trust-me benchmarks. A vendor says their model scored above human experts on something. And you either believe the benchmark or you don’t. There is no way to independently verify it without redoing the evaluation, which almost nobody does.

This result is different.

The counterexample is a specific formula, not a benchmark score. You can paste the polynomial map into SageMath, Mathematica, Wolfram Alpha, or any computer algebra system and confirm three things in under five minutes:

1. The Jacobian determinant is -2 everywhere. Not approximately. Exactly.
2. The three points listed above all produce the same output. Not close. Identical.
3. The polynomial has degree 7, not 100.

That verifiability is why this result spread so fast through the math community.

Researchers who normally roll their eyes at “AI solves famous problem” headlines went quiet when they could check the claim themselves in real time. Hacker News hit the front page within a day. Multiple news outlets published within 24 hours. When was the last time a result in algebraic geometry trended on social media?

If you ship AI products for clients, the lesson transfers directly. When you claim your AI did something impressive, attach the proof. Show inputs and outputs. Make it reproducible. If a working mathematician can verify a claim about algebraic geometry in five minutes using free tools, your case study about productivity gains can include actual before-and-after numbers.

What Does Claude Fable 5 Mean for AI-Assisted Research?

The two-dimensional case of the Jacobian conjecture remains open. The result has not been peer-reviewed as of July 20, 2026. It is possible, though unlikely given the simplicity of the check, that some subtlety will emerge during formal review.

Treat the “disproved” framing with appropriate caution.

But the strategic signal is already clear regardless of how peer review shakes out.

Frontier AI has crossed from competition math into genuine research collaboration. Not “AI wrote a paper that sounds like research.” Claude Fable 5 helped find a specific mathematical object that resolves a question that stood open for 87 years and sat on a list next to the Riemann Hypothesis. The human framed the problem. The AI proposed candidates. The human verified independently. That pipeline works today, right now, for problems a lot more practical than algebraic geometry.

If you run a small operation or build products solo, think about what that pipeline looks like in your domain.

You do not need a math PhD to benefit from this pattern. You need a hard problem, the ability to frame it precisely for an AI. And the discipline to verify the output yourself. The mathematician did not trust Claude Fable 5’s claim that the counterexample worked. He checked it with Wolfram Alpha. The checking is the work.

The workflows that produce real value share the same structure: human poses the question, AI generates candidates, human verifies.

The workflows that waste time skip the third step. This result is the cleanest example I have seen of why that third step matters. And why skipping it is the most expensive shortcut you can take in AI-assisted work.

Claude Fable 5 spent its early weeks embroiled in export-control controversy before being restored to a limited set of approved organizations. Now it has a math result that gives a better headline than regulatory chaos. If you are betting your stack on a model provider, watch how fast narratives flip. A few weeks ago the story was “too dangerous to ship.” This week it is “too good to ignore.”

The next time someone tells you AI is just autocomplete with confidence, send them the Wikipedia page for the Jacobian conjecture. It updated last weekend — and the tool that did it was Claude, not a roomful of mathematicians.