What Fable 5.2 showed

The claims

Anthropic released Claude Fable 5.2 on November 9, 2026. Its announcement leads with four results. Two open problems in black-hole physics were solved. Fermat’s Last Theorem was formalized in 30 hours. Three electronic devices were designed from a one-paragraph prompt. A simulated robot was controlled far better than before.

The announcement runs them together. They are different kinds of claim, and each has its own test. All four are reported by Anthropic.

Physics: a long calculation, checked by the same family of model

Both results rest on holographic duality, the idea that certain universes containing gravity can be described exactly by a quantum theory with no gravity. Both are preprints on the Hexagon server, labelled primarily AI-generated text, and neither is peer reviewed.

The first asks whether a black hole still forms through a sharp transition when the quantum theory is weakly coupled. The answer turns on the sign of one number, which the authors say had not been computed in the two decades since it was posed. They report that the transition is sharp. This is a calculation, a long one with a definite answer. Its check is a physicist testing special cases and a second Claude computation by a different method, which is not published.

The second finds the exact limit of a technique physicists have used for more than 15 years to rule out candidate theories of quantum gravity, a bound on the lightest excitation in a family of two-dimensional theories. The limit was known only to lie between 8.503 and 12. The paper shows it is 9.1266, from a closed formula. A physicist checked every main result, and the public Opus 5.5 model later reached the same value on its own. This is the stronger of the two.

The announcement credits Fable 5.2. The preprints credit “an internal Claude research model”.

Fermat’s Last Theorem: the headline number has a footnote

Andrew Wiles proved the theorem in 1995. What is new is checking the proof line by line in the Lean proof language. That is verification, not new mathematics, and Anthropic says so.

In September an unreleased Claude research model took 11 days to produce the first complete formalization, with a purpose-built harness and guidance from a human researcher. In the Fable 5.2 run about 1,000 agents did it in 30 hours, with a third of the action tokens and a proof seven times shorter.

On November 11, two days after release, Anthropic amended the announcement. The agents were given a list of 60 statements, a one-sentence sketch of the classical route, the definitions those statements use, and 47 proved lemmas from the first formalization. So 30 hours is the second pass, started from the first pass’s structure. It is not 30 hours against 11 days from the same start.

The comparison with Opus 5.5 is one run each. In the same set-up Opus 5.5 completed 26% in 18 hours and barely moved after that. What stands is the part that needs no trust: Lean accepts or rejects every step, so the agents’ word is not what is relied on.

What is different from Opus 5.5

Less than the announcement suggests, and in one direction. Epoch AI’s index, which is independent, puts Fable 5.2 first of 240 models at 173. Gemini 4 Carbon is at 170, Opus 5.5 and GPT-6 Astra at 169.

Anthropic’s own table has Fable 5.2 ahead of Opus 5.5 on most tests. On FrontierCode 2.0, which asks whether a model’s code changes would be merged by the maintainers of open-source projects, it scores 49.1% against 45.5%. On Humanity’s Last Exam with tools, 74.2% against 67.7%. On BenchCAD 2.1, hardware design, 83.5 against 65.3.

The gap is in how long a team of agents holds together on one task. Fable 5.2 coordinated more than 1,000 subagents for eight days to build an operating system that boots on a laptop. Anthropic’s account of the first Fermat attempts is that the agents “quickly lost track of the project’s state” until the work was kept as a graph of statements. The announcement describes a model that stays coherent for longer, not a new method.

What has not been shown

Outside Anthropic, the only measurement is Epoch’s index. We found no published review of the physics by physicists outside the company, no second run of the Fermat formalization, and no independent robot score for Fable 5.2. The one independent robot study tested a different model (Fable 5.2 and the world-model question).

What it means for our work

Everything Fable 5.2 did well has a checker: Lean, a second calculation, an evaluator that scores circuits. Whether an idea is right for one person has one checker, that person. So the gains help with the half of our work a machine can check, making and developing candidates, and leave the other half where it was, choosing among them. A strong model asked which New Yorker caption was funnier agreed with the crowd 71% of the time, against 69% for one random voter, and 51% against 57% among the top few captions (A model judge is one voter).

A better generator also moves the baseline. The plain answer a model writes with no framework has just improved, and what a framework adds is measured against that plain answer, so every stronger model raises the bar.

The studio is the more exposed half. Anthropic had Fable 5.2 design a home voice hub, a USB speakerphone and a handheld virtual pet from one paragraph, with the boards routed, the enclosures modelled and the software written, and all three worked. Geometry and drawing are the work general models are absorbing fastest. What stays the lab’s is which design is made, for whom, and why that person would keep it.

On that last point, public joke ratings show a selector that learned one person’s picks beating a shared judge by 2.9 points after 20 ratings and 6.4 after 80 (A selector that learns one person). Ideas that are new and right for the person asking sit in the part a stronger model leaves open.