On November 17, 2026, Anthropic said its Claude Fable 5.2 model had solved two open problems related to black-hole physics. The announcement describes a long-running calculation in quantum field theory and a new bound in the modular bootstrap, a method physicists use to rule out inconsistent theories. The two papers are public preprints, and both list Claude (Anthropic) as the author.[1][2][3]
The headline is easy to repeat. The harder question is what “solved” means here. Did a model produce a plausible answer? Did other researchers check it? Can another person follow the proof? The two projects have different answers, and that difference is more useful than treating them as one general claim about artificial intelligence.
This article compares the announcement with the papers themselves. It is a source-based explanation, not an independent review of the physics or a claim of hands-on replication.
The two problems are not the same
The first paper studies the Hawking-Page transition in a particular quantum theory. In simple terms, it asks whether a mathematical model of a hot universe changes abruptly from a particle-filled state to a black-hole state as its temperature rises. The calculation concerns N=4 super Yang-Mills theory, a quantum field theory with four supersymmetries. It is defined on a three-dimensional sphere. The paper studies weak coupling, where the interactions are small. It also takes the large-N limit, treating the number of color degrees of freedom as large. These conditions define the model. They do not describe a laboratory experiment or our universe directly.[2][4]
The second paper studies the spinless modular bootstrap in a two-dimensional conformal field theory. A conformal field theory is a quantum theory with symmetry under changes of scale and angle. The bootstrap approach applies consistency rules to narrow the possible theories. The paper asks how strongly one version of the method can bound the dimension of the lightest non-vacuum state as the central charge, a parameter that measures the theory's degrees of freedom, becomes large. It finds an exact limiting slope for a defined class of bounds.[3]
Both projects connect to black-hole questions through theoretical physics. Neither paper reports detecting a black hole, finding a new object in space, or resolving quantum gravity as a whole. Their results apply to specific mathematical models and stated assumptions.[1][2][3]
Result one: a transition at weak coupling
The Hawking-Page transition is a change between two dominant descriptions of a system in anti-de Sitter space, a curved space used in theoretical models of gravity. At lower temperatures, thermal particles dominate. At higher temperatures, a black-hole solution dominates. The quantum theory linked to that gravity model gives physicists another way to study the question.
Earlier work had identified a difficult three-loop calculation, meaning a calculation that includes third-order corrections, whose result would determine whether the weak-coupling transition is first order. A first-order transition is an abrupt switch between phases. The new paper computes the sign of a quartic coefficient, the number multiplying a fourth-power term, in the effective action for the Polyakov loop, a quantity used to describe the transition. The authors report the coefficient as b = -1.48 × 10^-3 λ², where λ is the theory's coupling strength. Under the paper's assumptions, that negative sign implies a first-order deconfinement transition, an abrupt change from a confined to a deconfined phase, at weak coupling and large N.[2]
The paper says the calculation was carried out by an AI system and describes checks used to reduce the chance of an arithmetic or implementation error. Its calculations were organized so that almost every number was obtained in two independent ways, or checked against an exact result. It also compares parts of its calculations with known results. For example, its methods reproduce earlier two-loop results in restricted cases. One comparison with published results differed. The authors say the difference was too small to change the sign of the coefficient that supports their conclusion.[2]
These checks matter, but they should be described precisely. Two independently written programs agreeing is stronger than one program repeating itself. Agreement with an established result tests parts of the calculation. Neither check alone proves that the model, assumptions, and interpretation answer every related physics question.
The paper states an important limit directly: it does not show that the first-order transition continues without interruption from weak to strong coupling. Its result establishes what happens near weak coupling, at leading order in 1/N, and through order λ². The larger question of how the transition behaves at intermediate coupling remains open.[2]
Result two: a sharper bootstrap bound
The second result is about the ceiling produced by one version of the modular bootstrap. In a two-dimensional conformal field theory, modular invariance relates descriptions of the theory at different temperatures. Researchers can apply carefully chosen mathematical functionals to that relation and derive bounds on the lowest scaling dimension above the vacuum.
The paper reports that the best large-central-charge slope for the spinless method is approximately 9.1266. A larger slope gives a stronger upper bound. The value 12 would reach the threshold associated with BTZ black holes, so 9.1266 is a sharper limit than earlier proved values, but it does not reach that threshold.[3]
The result has two parts. The paper argues that no functional in its specified class can improve the asymptotic slope beyond the value it derives. It also constructs an explicit functional that reaches the bound to the stated order. The proof uses mathematical analysis and computer-assisted estimates. The paper describes certified numerical calculations, including interval arithmetic, where a computer bounds numerical error rather than merely printing a decimal approximation.[3]
Anthropic adds a separate verification detail: it says the results were formally verified in Lean in September 2026, and that a physicist checked every main result before publication. Lean is a proof assistant, software that checks whether a formal proof follows from stated definitions and rules. This is meaningful evidence for the formalized claims. It is not the same as saying that Lean independently chose the research question, confirmed every physical interpretation, or verified every sentence in the paper.[1]
The paper also makes clear that a mathematical bound on a method is not a construction of the correct physical theory. It constrains a class of possible spectra under the assumptions in the paper. It does not identify which theory describes nature. The result is about the power of this bootstrap setup, not a final theory of quantum gravity.[3]
How to read the verification claims
When an AI research announcement uses the word “solved,” it helps to separate four questions.
What exactly is the claim? In these papers, one result is a sign calculation with consequences for a weak-coupling phase transition. The other is an optimality result for a defined class of mathematical bounds. Those are concrete statements. “Claude solved black holes” would be much broader than either paper supports.[1][2][3]
What was checked independently? For the Hawking-Page calculation, the paper reports independent computations and comparisons with known results. Anthropic says a researcher using Claude independently reached the same result by a different method. Agreement across different routes is useful, although both routes still need clear assumptions and inspectable derivations.[1][2]
What does formal verification cover? Anthropic says the modular-bootstrap result was formalized in Lean. A proof assistant can check a proof after the relevant definitions, lemmas, and statements have been encoded. That check does not automatically establish that the formal statement matches the intended physical question. Readers still need to understand the translation from the research claim to the formal theorem.[1][3]
What remains outside the result? The first paper does not settle the behavior between weak and strong coupling. The second does not identify the correct theory of quantum gravity. These limits are part of the results, not footnotes to ignore.[2][3]
This way of reading is useful beyond physics. For any AI-generated research, ask for the exact theorem or measurement, the assumptions, the verification method, and the authors' stated limits. A benchmark score or a confident announcement is not a substitute for those details.
What the papers show about AI-assisted research
The papers offer a case study in a workflow that combines model-generated work with human direction and checking. Anthropic's account says researchers set up agents to explore candidate problems. For the Hawking-Page result, the company says a physicist who had independently used Claude on the problem shared a result, and a colleague then recognized that an internal agent had reached the same answer earlier. The two approaches used different methods. For the bootstrap result, the company says multiple agents contributed, two other agents independently reached the same answer, and a physicist checked each main result.[1]
These details come from Anthropic's account of its own research process. They are not an outside audit of the full development history. The papers give readers a more direct basis for examining the mathematical claims, while the announcement describes how the company says the work was produced. Keeping those source types separate avoids treating a company account as independent confirmation.
The strongest lesson is not that an AI model has replaced a physics team. The work described includes problem selection, prior results, multiple computation paths, a proof assistant for one result, and human checking. A model may have produced important steps, but confidence should follow the visible verification, not the model's name or the scale of the headline.
The announcement shares both papers as preprints, making their arguments available for scrutiny. Readers can inspect the calculations and proofs, test the stated assumptions, and assess whether the results hold up. That review is still part of the story.
Sources
- Anthropic, “Claude Fable 5.2 solves two open problems in the physics of black holes,” November 17, 2026
- Claude (Anthropic), “The Hawking-Page transition at weak coupling,” preprint
- Claude (Anthropic), “The spinless modular bootstrap at large central charge,” preprint
- Ofer Aharony et al., “The Hagedorn/deconfinement phase transition in weakly coupled large N gauge theories,” arXiv
Top comments (0)