Claude Fable 5 Solves an 87 Year Old Math Mystery With a Tiny Formula

Something happened last week that the mathematics community is still processing. An AI model produced a short string of polynomial notation, roughly 216 characters long, and in doing so dismantled a conjecture that some of the brightest minds in algebraic geometry had been circling for nearly nine decades. The Jacobian Conjecture, open since 1939, appears to have fallen. Not to a tenured professor laboring over it for years. Not to a collaborative proof spanning hundreds of pages. To a formula spit out by Claude Fable 5, Anthropic’s latest reasoning model, and then verified by hand by a human mathematician who presumably had to sit down afterward.

This is not a story about AI getting better at math competitions. This is something categorically different.

What the Jacobian Conjecture Actually Claims

For those outside pure mathematics, the conjecture is deceptively simple to state. It concerns polynomial maps in several complex variables and asks whether a specific condition on their derivatives (the Jacobian determinant being a nonzero constant) guarantees that the map can be reversed. In plain terms: if a polynomial function never “collapses” locally, does that mean it never collapses globally? Mathematicians believed the answer was yes. For 87 years, nobody could prove it, and nobody could disprove it.

What Claude Fable 5 did was find a concrete counterexample. The model constructed a polynomial map in three complex dimensions whose Jacobian determinant equals negative two (a nonzero constant, satisfying the conjecture’s condition) but which sends three distinct input points in C³ to the same output. That is a direct violation. The map looks invertible everywhere locally but is not injective globally. The conjecture is false.

Levent Alpöge, a mathematician whose work spans analytic number theory and arithmetic geometry, verified the calculations by hand before the result went public. That human verification step matters enormously here, and we will return to it.

Why This Is Not Just Another AI Benchmark Win

The AI industry has grown accustomed to a certain rhythm of announcements. A model scores higher on a math olympiad. A model passes a bar exam. A model writes better code on HumanEval. These benchmarks, while useful for tracking progress, measure performance on problems that humans have already solved. They test whether AI can replicate known reasoning paths faster or more reliably.

The Jacobian Conjecture counterexample is fundamentally different. This was an open problem. There was no known solution for the model to pattern match against. No training example in any dataset contained the answer because the answer did not exist in human knowledge until the model produced it.

That distinction puts this result in a vanishingly small category. DeepMind’s AlphaFold predicted protein structures that experimentalists later confirmed. DeepMind’s FunSearch discovered new mathematical constructions in extremal combinatorics last year. But disproving a major named conjecture that has resisted almost a century of human effort occupies a different tier of achievement entirely.

The brevity of the solution makes it even more striking. A 216 character formula. The entire counterexample fits in a tweet. Mathematics has seen this before with human discoveries. Occasionally the answer to a deep question turns out to be shockingly compact, hiding in plain sight. But it took a machine to see it, and that fact will reshape conversations about the relationship between AI and mathematical research for years to come.

The Verification Problem and Why It Cuts Both Ways

Alpöge’s manual verification is the linchpin of this story’s credibility. Without it, the claim would be interesting but unconfirmed. With it, the mathematical community has a concrete, checkable object to scrutinize.

This points to something important about how AI might reshape mathematics going forward. Verification of a counterexample is dramatically easier than discovery. Checking that three specific points map to the same output under a given polynomial is undergraduate level computation. Finding those points and that polynomial in an infinite search space is the hard part. AI handled the creative search. A human handled the confirmation. The division of labor here is almost perfectly efficient.

But this dynamic also introduces a new kind of risk. As AI models produce increasingly complex mathematical claims, not all of them will be as easy to verify as a polynomial counterexample. Proofs that span novel techniques, introduce new definitions, or rely on extensive case analysis will be far harder for human reviewers to check. The mathematical community will need to develop new verification infrastructure, likely involving a combination of formal proof assistants like Lean or Coq alongside human review.

There is also the uncomfortable question of trust. If Alpöge had not verified the result, how long would the community have debated its validity? What happens when the next AI generated result lacks a willing and capable human verifier? The peer review system in mathematics already strains under the volume of human generated papers. Adding machine generated conjectures and proofs to the pipeline without scaling verification capacity could create a bottleneck that actually slows progress rather than accelerating it.

What This Tells Us About Anthropic’s Position

Anthropic has been deliberate in positioning itself as the safety focused AI lab, sometimes to the frustration of observers who wanted to see it compete more aggressively on raw capability. Claude Fable 5 changes that narrative. This result demonstrates frontier capability in a domain that carries enormous prestige and relatively low controversy. Pure mathematics is perhaps the safest possible arena in which to showcase a breakthrough. There are no deepfakes, no misinformation risks, no job displacement anxieties. Just a clean, verifiable intellectual achievement.

The timing is also notable. OpenAI recently released o3 and o4-mini with strong mathematical reasoning. Google DeepMind continues to publish results on AI for mathematics and science. xAI’s Grok models have been pushing into technical domains. Anthropic needed a flagship demonstration that the Claude model family can do more than write careful, well hedged text. Disproving a named conjecture from 1939 is about as flagship as it gets.

Whether this translates into commercial advantage is a different question. Enterprise customers evaluating Claude for coding assistance or document analysis are unlikely to change procurement decisions because the model solved an algebraic geometry problem. But it signals depth of reasoning capability that matters for the hardest technical use cases in finance, drug discovery, materials science, and engineering. And for the research community, which influences long term talent pipelines and institutional partnerships, this result carries real weight.

The Broader Arc of AI in Mathematics

It is worth placing this moment in a longer timeline. For decades, computers assisted mathematicians primarily through brute force computation. The four color theorem proof in 1976 relied on exhaustive case checking by machine. More recently, AI has moved from computation to conjecture. DeepMind’s collaboration with mathematicians in 2021 used machine learning to spot patterns in knot theory and representation theory, leading to new human driven proofs. FunSearch took another step, generating novel mathematical objects through large language model guided search.

The Jacobian Conjecture result goes further still. The model did not spot a pattern for a human to formalize. It did not assist a proof. It produced the complete counterexample. The creative act, such as it is, was the machine’s. The human role was verification.

This progression from computation to pattern recognition to conjecture assistance to independent discovery traces an unmistakable trajectory. Within the next few years, it is reasonable to expect AI models to routinely generate novel mathematical results, some of which will be significant. The bottleneck will shift decisively from discovery to verification, and the institutions that adapt to this shift fastest will have an outsized influence on the direction of mathematical research.

What People Are Overlooking

Most commentary has focused on the achievement itself. Less discussed is what it implies about the model’s internal representations. Claude Fable 5 did not solve this problem by searching through polynomial maps at random. The search space is effectively infinite. The model had to develop or leverage some form of structural understanding of polynomial maps, Jacobian determinants, and injectivity in several complex variables. What that understanding looks like internally remains opaque, and that opacity is itself a significant issue.

If AI models are developing mathematical intuition that humans cannot inspect or interpret, the field faces a novel epistemological situation. Mathematics has always been grounded in the principle that any result can be checked by any sufficiently trained person. When the discovery process becomes opaque but the result remains checkable, the social contract of mathematics changes in subtle ways. Mathematicians may increasingly find themselves in the position of verifying claims they did not originate and cannot fully explain.

There is also a competitive dynamic that deserves attention. If one AI lab’s model can disprove longstanding conjectures, funding agencies and university departments will face pressure to provide access to frontier models as standard research infrastructure. This could accelerate a shift where mathematical progress becomes partially dependent on commercial AI providers, raising questions about access, reproducibility, and the concentration of intellectual capital.

What Happens Next

The Jacobian Conjecture counterexample will undergo intense scrutiny. Other mathematicians will verify Alpöge’s verification independently. If the result holds, textbooks will need revision, and several related lines of research will require reevaluation. The conjecture is connected to other open problems in algebraic geometry, and its resolution (in the negative) will have downstream consequences that mathematicians will spend years working through.

For the AI industry, this result will intensify the race to demonstrate scientific discovery capabilities. Expect OpenAI, Google DeepMind, and others to highlight their own models’ mathematical achievements with renewed urgency. The narrative of AI as a tool for genuine discovery, not just productivity enhancement, is now anchored by a concrete and dramatic example.

For anyone building on top of AI reasoning capabilities, the lesson is that the ceiling on what these models can do in structured, formal domains is higher than most people assumed even six months ago. A 216 character formula just rewrote an 87 year old chapter of mathematics. The next chapter is being written right now, and the authors are no longer exclusively human.

For nearly nine decades, the Jacobian conjecture sat comfortably among the most stubborn unsolved problems in algebraic geometry. First posed by Ott-Heinrich Keller in 1939, the question seemed almost deceptively simple: if a polynomial map has a nonzero constant Jacobian determinant, must it be globally invertible across complex Euclidean space? Generations of mathematicians attacked it from every conceivable angle. None succeeded. Then Anthropic’s Claude Fable 5, working alongside number theorist Levent Alpöge, produced a one line counterexample that fits in a tweet.

The result, if it survives peer review, will rank among the most consequential AI contributions to pure mathematics to date. It highlights the growing intersection between AI-driven discoveries and the oldest and most abstract of human intellectual pursuits.

AI didn’t just solve a math problem — it forced mathematics to reckon with a new kind of intellectual partner.

What Actually Happened

Alpöge, a researcher at Anthropic, reportedly posed the Jacobian conjecture to Claude Fable 5 during what he described as a casual session that happened to coincide with a FIFA World Cup final broadcast. The AI system responded with a specific polynomial map from complex three dimensional space to itself. The formula spans roughly 216 characters.

The construction is elegant in its simplicity. The polynomial map carries a Jacobian determinant equal to the constant value of negative two everywhere, satisfying the conjecture’s hypothesis perfectly. Yet the map sends three distinct input points in C³ to the same output point, which directly proves non injectivity. You cannot have a globally invertible function that maps multiple inputs to the same output. The conjecture, at least in three or more dimensions, is dead.

Alpöge verified every calculation by hand before sharing the explicit counterexample in a brief post on X. That detail matters enormously. This was not a case of an AI system declaring victory and everyone nodding along. A trained mathematician checked each step manually, and the formula is compact enough that any algebraist can do the same.

Why This Is Different From Previous AI Math Results

The temptation is to lump this in with earlier headline grabbing moments where AI systems tackled mathematical problems. Google DeepMind’s AlphaGeometry solving International Mathematical Olympiad problems last year generated significant attention, as did earlier work by systems like AlphaProof. Those achievements were real, but they operated within relatively bounded domains where the search space, while large, was well characterized.

The Jacobian conjecture is a different animal entirely. This is a problem that sat on the Clay Mathematics Institute’s radar and appeared on various lists of important open problems in algebra. The difficulty was not computational brute force but rather the creative construction of an object that nobody could find. The polynomial’s coefficients and structure had to be carefully engineered to maintain a constant determinant while simultaneously funneling multiple inputs to a single output. That kind of construction has a creative dimension that goes beyond pattern matching or search optimization.

What Claude Fable 5 appears to have done is navigate an enormous space of possible polynomial maps and identify one with precisely the right properties. Whether this reflects genuine mathematical reasoning or an extraordinarily effective form of guided search is a question that will occupy AI researchers and philosophers of mathematics for years. But from a practical standpoint, the distinction may matter less than the outcome.

The Mathematical Fallout

The counterexample’s existence in dimension three cascades upward immediately. Standard results in algebraic geometry mean the conjecture now fails in every dimension three and above. The entire edifice of proof strategies that mathematicians had been developing for higher dimensional cases collapses overnight.

However, the two variable case remains completely open. The original conjecture for polynomial maps from C² to C² is untouched by this result, and it continues to stand as a central unsolved problem. This creates a fascinating bifurcation in the field. Researchers who had been working on the general case must now redirect their efforts toward classifying which polynomial maps with constant Jacobian determinants are invertible and which are not. The problem has shifted from “prove or disprove” to “map the boundary between true and false.”

That pivot will likely generate an entirely new subfield of research. And it raises an obvious question: could Claude Fable 5 or a successor model eventually settle the two dimensional case as well? The roots of this problem stretch back even further than Keller, since Ludwig Kraus first stated the conjecture for two dimensions in 1884.

What This Tells Us About Anthropic’s Strategy

Anthropic has been notably more restrained than OpenAI or Google in making dramatic public claims about its models’ capabilities. The company has focused its public messaging on safety research, constitutional AI, and responsible scaling. Having a landmark mathematical result emerge from what was apparently an informal session rather than a staged demonstration fits that pattern. There was no press conference, no embargo, no carefully choreographed reveal. Alpöge posted it on X.

But make no mistake about the strategic significance. Anthropic has been locked in an increasingly intense competition with OpenAI, Google DeepMind, and now xAI for talent, compute, and credibility. A result like this does more for Anthropic’s reputation among the scientific community than any benchmark score ever could. It signals that Claude models are not just commercially useful chatbots but tools capable of contributing to frontier research.

The fact that Alpöge is both a number theorist and an Anthropic researcher is itself telling. Anthropic has been quietly building a bench of researchers who bridge domain expertise and AI development. This approach mirrors what Google DeepMind did with AlphaFold, where deep biological expertise within the team was essential to framing the problem correctly for the AI system.

The Human AI Collaboration Model

The most important structural detail in this story is the division of labor. Claude Fable 5 handled the combinatorial and algebraic search for a viable counterexample. Alpöge supplied the mathematical framing, posed the right question, and performed independent verification. Neither could have done this alone, at least not in the same timeframe.

This is the collaboration model that AI optimists have been describing for years, but concrete examples at this level of significance have been vanishingly rare. Most real world AI assisted research has involved tasks like protein structure prediction, materials discovery, or code generation, domains where the AI’s contribution, while valuable, operates within relatively well defined boundaries.

Resolving an 87 year old conjecture is qualitatively different. It suggests that frontier AI models are beginning to function as genuine research partners in pure mathematics, not just in applied or computational domains. If this pattern holds, the implications for how mathematical research is conducted over the next decade are profound.

What People Are Overlooking

Three things deserve more attention than they are getting.

First, the verification bottleneck. The counterexample is short enough for hand verification, which is why confidence in the result is already high even before formal peer review. But most mathematical results are not this compact. As AI systems tackle increasingly complex problems, the gap between AI generated proofs and human verifiability will widen. The field urgently needs robust formal verification systems that can bridge that gap, and companies like Anthropic and Google DeepMind know it.

Second, the training data question. Claude Fable 5 was trained on vast amounts of mathematical text, including decades of published work on the Jacobian conjecture. To what extent did the model synthesize existing partial results and intuitions scattered across the literature versus generate something genuinely novel? The answer has significant implications for how we credit AI contributions and how we think about mathematical originality.

Third, the incentive structure for mathematicians. If AI systems can resolve longstanding conjectures, what happens to the career incentives that drive human mathematicians to spend years on a single problem? This is not an abstract concern. Young researchers choosing specializations will factor in the likelihood that an AI system might solve their target problem before they do. The field may see a shift toward problems that require the kind of conceptual framing and question posing that AI systems still struggle with, while computational and constructive problems increasingly become AI territory.

The Peer Review Question

Formal publication and peer review are underway, and this is where the story could still take a turn. The mathematical community has been burned before by claimed proofs that collapsed under scrutiny. The compact nature of the counterexample works in its favor, since verification is straightforward. But the review process will also examine whether the polynomial map truly satisfies all necessary conditions in full generality, not just for the specific test cases that have been checked so far.

The early signs are encouraging. Alpöge’s reputation as a number theorist is strong, the formula is publicly available for anyone to check, and no errors have been identified in the days since the initial post. But until the result passes formal peer review and is accepted by a major journal, it remains a preprint level claim, however compelling.

Where This Goes Next

If the proof holds, expect a cascade of follow on effects.

Other AI labs will accelerate their own efforts to apply frontier models to open mathematical problems. OpenAI, which has invested heavily in reasoning capabilities with its o series models, will face pressure to demonstrate comparable results. Google DeepMind, already active in this space through AlphaGeometry and AlphaProof, will likely intensify its own efforts.

Funding agencies and research institutions will begin grappling with how to evaluate grant proposals that involve AI collaboration. Should a mathematician who uses an AI system to find a counterexample receive the same credit as one who constructs it through traditional methods? The academic incentive system was not designed for this.

And the broader AI industry will have a new, concrete example to point to when discussing the value of frontier model development. This is not a chatbot answering customer service queries or generating marketing copy. This is a contribution to human knowledge that stands on its own merits, verified by human experts, enabled by machine intelligence.

The Jacobian conjecture survived 87 years of human effort. It did not survive its first serious encounter with a frontier AI model and a mathematician who knew exactly the right question to ask. That combination, human insight directing machine capability with rigorous verification closing the loop, may turn out to be the defining paradigm of the next era of scientific discovery.