Frontier Math × AI
Claude Fable 5 dropped
an 87-year-old assumption.
The Jacobian Conjecture — a 1939 claim in algebraic geometry that most people quietly assumed was true — just took its first serious hit. On July 20, 2026, Anthropic researcher Levent Alpoge reported that, working with Claude Fable 5, he constructed a concrete counterexample: a polynomial map $\mathbb{C}^3 \to \mathbb{C}^3$ whose Jacobian determinant is a nonzero constant, yet with no polynomial inverse. Mathematicians verified it by hand within a day.
The Announcement
"Probably true" got
broken by one formula
The Jacobian conjecture, posed by Ott-Heinrich Keller in 1939, claims that if a polynomial map $F: \mathbb{C}^n \to \mathbb{C}^n$ has a nonzero constant Jacobian determinant, then $F$ has a polynomial inverse defined globally. ForkLog and Gigazine's English edition reported that on July 20, 2026, Levent Alpoge, working with Claude Fable 5, produced an explicit counterexample in $n=3$.
A counterexample isn't a proof; it's a single hole in what the conjecture claims must be true. Alpoge's polynomial map $F: \mathbb{C}^3 \to \mathbb{C}^3$ satisfies the premise — its Jacobian determinant is $-2$, nonzero and constant — yet it has no global polynomial inverse. One counterexample in any dimension is enough: the "true for all $n$" claim collapses. As Kingy.ai emphasized, the game changed from "produce a proof" to "produce one witness object," which is a very different shape of problem.
| Jacobian conjecture, until now | The Alpoge × Fable 5 counterexample |
|---|---|
| 87 years of failed proof attempts | One witness collapses the universal claim |
| Named by Smale on "Problems for the 21st Century" | Focus shifts to why the partial results still hold |
| Computer searches never produced a counterexample | AI narrows the search; humans verify by hand |
Instead of a proof,
bring back a single witness that breaks the claim.
Why This Format Matters
The counterexample
is instantly checkable
The strongest place AI can help mathematics is not proofs, but producing discrete objects whose validity is decidable by substitution. This is a textbook case.
Claude Fable 5 generated candidates
Alpoge fed the model instructions of the form "look for a Jacobian counterexample with this structure" and iterated. Rather than brute-forcing the space numerically, the leverage was in generating shapes that could be a counterexample — a place where LLMs have real value.
Alpoge did the filtering
The judgment of "does this candidate really have a nonzero constant Jacobian determinant and lack a polynomial inverse" was the mathematician's, not the model's. This success is a case study in AI candidate generation × human suitability judgment, structured as a working dialogue loop.
Verified in under 24 hours
The counterexample is one explicit polynomial map — a discrete object whose validity is decidable by substitution. Other mathematicians confirmed it by hand within 24 hours, as ExplainX.ai reported. Formal peer review remains, but that speed is unheard of for anything shaped like a proof.
Numbers to Sit With
Three numbers to hold
this result against
Don't overreach. The essence is the shape of the win: a witness object beats a claim.
Formal peer review is still ahead. As ExplainX.ai emphasized, the 24-hour verification is a first-pass substitution check; the deeper algebraic-geometric consistency review has yet to run. Until then, calling this a "very strong candidate counterexample" is closer to accurate than "the conjecture is refuted."
Even so, the case shows a concrete, repeatable playbook: if one discrete counterexample suffices, current AI plus one mathematician is enough. Proof-shaped attacks are still out of reach for AI; counterexample-shaped attacks are not. That's the useful line.
Who Should Care
Who this lands for
You don't need to care about Jacobian polynomials for the shape of this story to be useful elsewhere.
Researchers / Academia
Another proof that "AI-in-the-loop dialog to narrow candidates" is a real workflow. It looks especially strong for open problems on the counterexample side, not the proof side.
Engineers (search problems)
Any domain where one counterexample refutes a hypothesis — test design, security, SAT solving — is a natural fit for the same AI-in-the-loop search.
Product / Leadership
Another data point that frontier AI is producing citable research value. Useful ammunition when arguing for R&D-adjacent AI budgets.
Counter-view / Next
Don't overread this —
and what comes next
Three counter-points are due. First, "Claude solved the Jacobian conjecture" is wrong. Alpoge is a serious number theorist; the filtering and consistency judgments were his. Keep the AI-in-the-loop framing; "AI replaced the mathematician" is not what happened. Second, formal peer review is pending. History has cases where "counterexample" reports were retracted years later. Cautious language is right. Third, this method is strong when a claim can be refuted by a single witness. It does not port as-is to problems that require a proof (Riemann, P vs NP, etc.). Don't oversell the technique's reach.
Three near-term calls. (1) Expect more AI-in-the-loop counterexample-hunting teams across open conjectures in number theory, combinatorics, graph theory. (2) Anthropic will use this as a "Fable 5 contributed to a mathematical frontier" case study and likely ship tighter research-support tooling for math (specialized libraries and interfaces). (3) A norm shift may follow — attaching the AI dialogue log as an arXiv appendix could become common practice in math.
Three actions. (1) Any team with a research function should inventory their open problems and ask which can be reduced to a "witness object" shape. (2) Stand up a small PoC with a Fable 5 / Opus 4.8-class model as a candidate generator for one of them. (3) Get ahead of the credit question: settle your attribution rules for AI-in-the-loop results internally and, where possible, in your community norms. Better to decide before the next case forces the debate on you.