Fable 5.2 and the world-model question

The claim

With Claude Fable 5.2, released on November 9, 2026, the question is whether a language model has done what some said only world models could.

Yann LeCun’s version of the bet is narrow. Language models are great at language, code and math, he has said, but a world model lets a system forecast the consequences of its own actions and plan by search, and they do neither. Fei-Fei Li’s is broader. Language models, she wrote, are “wordsmiths in the dark; eloquent but inexperienced, knowledgeable but ungrounded.” Neither denies what the models do well. The dispute is about the physical world and about planning.

Why language goes so far

Language is the record of what people have learned about the world. Predicting it well means learning its structure. Geoffrey Hinton argues this goes as far as understanding. A model turns words into features, the features interact, and the derived features predict the next word’s features. That, he says, is what understanding is. At the AI4 conference in 2024 he said language models are the best model we have of how people understand language, and that the small language model he introduced in 1985 was a theory of how people learn what words mean.

The sciences moved their worlds into languages long ago. Galileo wrote in 1623 that the book of nature is written in the language of mathematics. Lavoisier opened his 1789 chemistry by citing Condillac, “the art of reasoning is nothing more than a language well arranged”, and argued that the language of a science cannot be improved without improving the science.

Biology borrowed its vocabulary the same way. Molecular biology took “code” and “information” from wartime cryptanalysis and linguistics between 1940 and 1967, in the historian Lily Kay’s account, and she argues the genetic code is a metaphor. When a field’s world is written in notation, a model fluent in the notation works in the field’s own medium. Lean, equations, code and circuit descriptions are such media.

An agent with tools gets its consequences from the world, not from a prediction. LeCun’s test is forecasting consequences and planning by search. An agent that writes code, runs it and tries again gets the consequences by trying. Lean accepts or rejects each step. An evaluator scores each circuit. A thousand agents searching for 30 hours are the search. The model of the world sits outside the weights, in the checker.

Anthropic’s protein result shows it plainly. Mythos 5.2, the same underlying model as Fable 5.2, did not learn capsid structure from text. It wrote code that ran two open-source structure-prediction models on overlapping groups of chains and merged the results, and Anthropic reports it modelled 28 of 34 capsids accurately. Quantum computing follows the pattern. In a preprint by authors at Harvard, QuEra and Anthropic, language models propose circuit designs and an independent evaluator verifies and scores them. In seven of eight tasks the circuits used fewer Toffoli gates and qubits than the state of the art.

The world model did not disappear. It moved into the tools, and the language model became the planner that drives them.

Where the premise breaks

Fable 5.2 is not language-only. It reads images, and the robot benchmark it was scored on is controlled from camera images. The claim that language alone did this is wrong twice: the model sees, and it leans on tools that model the world.

The wins are also not in the physical world. Proofs, physics calculations, code and circuits are worlds where a trial is cheap and the answer can be checked. The method stops working where a trial is slow, costly or physical.

The physical world is still open

Anthropic reports 64.9% for Fable 5.2 on RoboDojo Sim, a simulated two-arm robot controlled from camera images, against 39.4% for Opus 5.5 and 29.8% for GPT-6 Astra. Anthropic ran those scores itself.

The one independent study tested Astra. It measured 22.5% average success, above all 40 public robot policies it compared against, and weak results on tasks that need precision, dynamic control or two-arm coordination. Two other language models scored under 2% with the same set-up. The same model scored 29.8% in Anthropic’s run and 22.5% in the study, so only scores from one harness can be compared. A success rate shows a task was completed, not how the actions were chosen.

The other side is no further along. NVIDIA introduced GR00T 2 on November 9, a model that predicts future video and robot actions together. It reports 40% success zero-shot on 50 tasks never seen on the target robot, on NVIDIA’s own tasks. As of October 6, StartupHub reported that LeCun’s AMI Labs had released no public benchmarks or customer deployments. Neither camp has a decisive robot result.

What language is for

Yuval Noah Harari’s case for language is about power, not physics. “Language is the stuff almost all human culture is made of,” he wrote in The Economist in 2023. Human rights, money and gods are cultural artifacts made by telling stories and writing laws, so a system that masters language, he argued, has hacked the operating system of civilization. That is the strongest case that language is enough: much of the world runs on text, in law, money, code and contracts, and a model fluent in text can act on it without a body.

The counter is in the brain. Fedorenko, Piantadosi and Gibson reviewed the evidence in Nature in 2024 and concluded that language is primarily a tool for communication rather than thought. People who lose language can still solve math problems, play chess and plan. Others, Noam Chomsky among them, reject that view. If it holds, a model trained on language learns the part of thought that was made to pass between minds, and Fable 5.2 shows how much that part can do when something checks it. Li makes the same turn. She quotes Wittgenstein, “the limits of my language mean the limits of my world”, and answers that for AI there is more than words.

Language is also not only ours. Charles Hockett’s 1960 list of its design features includes displacement, speaking of things that are not here, and productivity, building endless new messages from parts. Honeybees dance the direction and distance of food away from the hive, work for which Karl von Frisch shared the 1973 Nobel Prize in Physiology or Medicine. A 2024 study of 8,719 sperm whale codas found that rhythm, tempo, rubato and ornamentation combine into about ten times as many distinguishable codas as had been thought. What they mean is unknown. In plain terms, language is a way of compressing what one mind knows so that another can use it, and other animals do a version of it.

The labs are moving together

• AMD agreed on September 28 to acquire World Labs for $8.2 billion, expected to close before the end of the year subject to regulatory approval. Li is to become AMD’s executive vice president and chief scientist.
• Google DeepMind’s Genie 4, introduced on October 27, builds on the Gemini family. Gemini Omni anchors each simulation in real-world appearance, and other Gemini models supply the reasoning that keeps the world coherent.
• NVIDIA calls GR00T 2 a world-action model.
• On September 3, ARC Prize said OpenAI’s GPT-6 Astra builds a compact symbolic world model to solve ARC-AGI-3 environments. Gary Marcus, who has argued for such models for years, called it “extraordinarily vindicating” and added that he suspects the best performance in verifiable domains.

Language models are taking on world-model jobs, and world-model labs are building on language models. “Language or world model” is the wrong question. The right one is where the model of the world lives, in the weights, in the tools or in a simulator, and what it costs to try.

What it means for our work

Our aim is ideas that are new and right for the person asking. In LeCun’s terms that needs a generator and a cost function for one person. The model of the world that matters to us is a model of a person, and the only one who can try an idea on that person is the person.

A framework is that model written in language, so our bet sits on the language side of this argument. With the name removed, a framework let a judge model identify its author 88% of the time against 25% by chance, mostly by carrying the person’s world rather than their method (A way of thinking travels). On public joke ratings, a selector that learned one person’s picks beat a shared judge by 2.9 points after 20 ratings and 6.4 after 80 (A selector that learns one person). Language carried a person far enough to be recognised and a little way toward being chosen for.

Fedorenko’s point bears on the design half. A framework carries what a person can say about how they decide. Much of design is not said, such as how a line sits or how a room takes light. We carry that in drawings, 3D models and renders, because words can describe a chair and only sitting tests it. A model of the physical world is the right tool for testing a design before it is built, and the world-model labs are building it. The generator can be a language model or a world model. What we add is the choice of what to build, and for whom (Who chooses what a world model builds). What Fable 5.2 changes in our own work is in What Fable 5.2 showed.