Can an AI agent, given a list of open problems and left to work, settle questions that physicists set aside for years? This post, published on 17 November 2026 under Anthropic's Science label with no individual byline, reports that Claude agents solved two open problems about black holes in model universes used to study quantum gravity. Its page heading is "Claude solves two open problems in the physics of black holes"; the page's title and description name the model: "Claude Fable 5.2 has solved two open problems in theoretical physics." The post explains both results for a general reader. The results themselves are in two preprints on Hexagon, a repository for research in the mathematical sciences and theoretical computer science, which are the versions of record for the physics:
- The Hawking-Page transition at weak coupling, hexagon:2611.00175v1 (42 pages);
- The spinless modular bootstrap at large central charge, hexagon:2611.00174v1 (131 pages).
Both are version 1 of 7 November 2026, credit "Claude (Anthropic)" as author, carry a CC BY 4.0 license, and are labeled on Hexagon "Primarily AI-generated text" and "Human understanding: some parts." Hexagon states that accepting a deposit "is not a judgment about its correctness, quality, or novelty." Neither paper is peer reviewed.
How the work was done
The post's section "A note about Claude's methods" describes the setup:
- The search. In August 2026 an Anthropic researcher, Hongbin Chen, "testing Claude's capabilities as a side project," asked Claude for open problems in theoretical physics focused on black holes and the bootstrap. Claude returned 22 candidates. Separate agents worked on each, could read each other's progress posts, and "ran in the background, with no special setup, for about six weeks." The post reports the two problems solved; it does not describe the outcome on the other twenty.
- The Hawking-Page problem. Weeks after an agent reached a solution, a physicist who had "independently used Claude to tackle the same problem" shared his result; the two used different methods. The preprint's provenance statement reads: "An internal Claude research model, prompted by a physicist, obtained the result presented here. The physicist checked it against several known special cases. A separate Claude computation using a different method, not included here, gives the same result." Neither document says which of the two computations the paper presents.
- The bootstrap problem. Several agents contributed: the one assigned the problem spotted a pattern in published numerical results, and another, working on a related sphere-packing problem, turned it into a method that reproduced a sphere-packing result "recently proved by an OpenAI model." An earlier agent had failed in "61 parallel attempts." Two other agents reached the same answer. The post says that "In September, the results were formally verified in Lean" and that "Claude then wrote the paper with minimal guidance, and a physicist checked every main result." The preprint's provenance statement says the result "was first obtained autonomously by an internal Claude research model," that "The publicly available Claude Opus 5.5 later obtained the same result independently," and that a physicist gave "brief guidance" and "checked every main result."
The post names Claude Fable 5.2 for the bootstrap result; the preprints say "an internal Claude research model" for both. Fable 5.2 was released on 9 November 2026, after the work described (see its system card). Neither document says whether the internal model was a Fable 5.2 snapshot.
The Hawking-Page transition at weak coupling
The question. In 1983 Hawking and Page showed that heating a universe with negative cosmological constant (anti-de Sitter space) produces an abrupt, first-order switch from a gas of particles to a black hole. Through the holographic duality, this corresponds to a deconfinement transition in N=4 super Yang-Mills theory on a sphere, but Hawking and Page's calculation holds only where gravity is classical, which is the theory's strong-coupling end. At weak coupling the theory also has a transition, found in 1999 and studied by Aharony, Marsano, Minwalla, Papadodimas and Van Raamsdonk (AMMPV). In 2003 AMMPV showed that whether it is first order depends on the sign of one number, the quartic coefficient $b$ in the effective action of the Polyakov loop, which first appears at three loops (preprint, Sections 1.2 to 1.5). They computed it in 2005 for pure Yang-Mills theory, from fourteen three-loop diagrams (negative, so first order), but never for N=4, where "The number of diagrams, species assignments and colour orderings is larger by more than a factor of ten" (Section 1.5). The post says the number "involves many thousands of terms, known as diagrams" ("The Hawking-Page transition"); the preprint gives no such count, and its own method is Hamiltonian rather than diagrammatic (Section 1.6).
The result. The preprint finds $b = -1.48 \times 10^{-3}\,\lambda^2$, with $\lambda$ the 't Hooft coupling (abstract), assembled from three pieces that are all negative; the three-loop piece, $-2.234$ in the paper's units, "would have to be +11.8 to change the sign" (Section 1.6, equation 6). The transition is therefore first order at weak coupling, lies slightly below the Hagedorn temperature, and its three saddle points "appear, exchange dominance and disappear in the same order as thermal anti-de Sitter space and the small and large black holes" (abstract).
The checks. The computation uses a Hamiltonian method rather than Feynman diagrams, and was organized so that "no number is used until two independent routes agree on it," with one exception: the coefficients of one series from its seventh order onward come from a single program (Section 6.6; Appendix C). Restricted to gluons, the same codes reproduce AMMPV's two-loop results to six digits. Of AMMPV's fourteen three-loop pure Yang-Mills diagrams, grouped into seven classes, five agree within rounding and a sixth within one unit in the last digit. The seventh does not: the preprint finds $-2.095$ against AMMPV's $-3.388$, "We have not resolved this," and "One of the two calculations has an error in this class." The disagreement is 0.8 in units of the three-loop term against a margin of fourteen for the sign of $b$, so the authors conclude it "does not affect our conclusion" (Section 8; Table 5). A check against the known two-loop Hagedorn temperature gives $-10.71(6)$ against $-10.7287$, but the paper states that "It is not a blind test": a first value of $-9.7$ led to the discovery of a missing counterterm, after which the numbers agreed (Section 8).
What is not shown. The paper does not show that the first-order line continues to strong coupling, works at leading order in $1/N$ and to order $\lambda^2$, and fixes one finite counterterm by fifteen exact conditions without an analytic derivation (Section 1.7). The post's statement that the transition is abrupt at both ends, so that "The most natural possibility is that it stays abrupt everywhere in between," is the post's reading. The preprint says only that a single first-order line from weak to strong coupling "is allowed," and that "Whether the line in fact continues unbroken through intermediate coupling is not something that a weak-coupling calculation can test" (Section 10.1).
The spinless modular bootstrap
The question. In 3D gravity in anti-de Sitter space, the bootstrap rules out candidate quantum theories by showing they are inconsistent. Hellerman's modular bootstrap (2009) uses the requirement that a torus partition function be unchanged when its space and time loops are swapped, and bounds the dimension $\Delta_1$ of the lightest state other than the vacuum, which the post describes as the mass of the lightest black hole (post; preprint, Section 1). At large central charge $c$, bounds of this kind read $\Delta_1 \le c/\Lambda + o(c)$, and a larger slope $\Lambda$ is a stronger bound. Before this work the best slope was known only to lie between 8.503, proved with analytic functionals by Hartman, Mazáč and Rastelli, and 12, where the bound would reach the black-hole threshold; numerical work by Afkhami-Jeddi, Hartman and Tajdini had extrapolated to about 9.08 and stated the estimate as $c/9.1$ (abstract; Section 1.2).
The result. For the spinless version of the bootstrap, in theories with $c > 1$ and no conserved currents beyond the Virasoro algebra, the preprint states a proof that the optimal slope is exactly
$$\Lambda_V = \frac{12\pi^2}{\pi^2 + 4\,\mathrm{arcsinh}^2 1} \approx 9.1266$$
(Theorem 1, abstract). The proof has two halves. No functional in a stated class, which contains every finite combination of derivatives, improves on $c/\Lambda_V$ by more than $O(c^{1/3}\log^{2/3} c)$ (optimality, Theorems 3 and 4, Section 6). An explicit functional gives $\Delta_1 < c/\Lambda_V + 0.483\,c^{1/3} + O(1)$ for all sufficiently large $c$ (attainment, Theorem 5 and Corollary 6, Sections 7 and 8). The attainment half is computer-assisted: it rests on finitely many inequalities certified in interval arithmetic (Sections 1.3 and 11.1; Appendix H). The proofs treat the logarithm of a functional as a two-dimensional electrostatic potential with charges at the functional's zeros and bound it by an explicit harmonic function (abstract; Section 6). A separate derivation of $\Lambda_V$ from the "spectral curve" of the optimal functionals, which the post describes as Claude's route to the result, is marked heuristic, rests on four named hypotheses, and is not used by the proofs (Sections 3 and 4).
Its scope. The rigorous statements assume $c \ge 40$ (Section 2), and the constants beyond which the asymptotic bounds hold are "explicit in principle, and we have not computed it" (Section 6). At every central charge reached by numerical work, the explicit bound is weaker than the existing numerical bounds (Section 8, discussing Corollary 6). A finer statement, that the best bound is $c/\Lambda_V + 0.27666\,c^{1/3} + O(1)$, is a conjecture supported by numerics up to $c = 2048$, which extrapolate to a slope of $9.1263 \pm 0.002$ (Conjecture 1, Sections 4.6 and 5). The post adds that this version of the method cannot by itself push the bound down to $c/12$, where black holes first appear.
Prior work it rests on. The sphere-packing result the post mentions is chapter 1 of an OpenAI document of August 2026, which the preprint cites for a proof of the slope $\pi^2$ in the related problem with $U(1)^c$ currents (a form of the Cohn-Elkies linear program for sphere packing, Section 2.3). The preprint says: "we have read the statements and the structure of that argument but have not independently verified it; nothing in this paper depends on it logically" (Section 1.2). The post says that the optimality half of the final proof "follows the strategy of the OpenAI proof."
The instructions given
Neither document prints a prompt. The post says the starting request asked Claude for open problems "specifically focusing on black holes and the bootstrap method," that the agents had "no special setup," and that the bootstrap paper was written "with minimal guidance." The Hawking-Page preprint's Appendix C records how the work was organized: many instances "each with a narrow brief," one coordinating through a design document, under three rules "imposed from the start" (no number enters unless two independently written programs agree or it matches an exact result; every program must pass a test case with a known answer before production use; "a disagreement is never averaged or argued away"). It does not say who imposed them.
Limitations
- Not peer reviewed, and AI-written. Both preprints are labeled primarily AI-generated, with human understanding of "some parts." Human checking is described as one physicist per paper: checks "against several known special cases" for the Hawking-Page result, and of "every main result" for the bootstrap.
- An unresolved disagreement. The Hawking-Page computation and AMMPV disagree in one class of pure Yang-Mills diagrams; the paper argues the difference is too small to change the sign of $b$ but does not say which side is wrong.
- A formal verification this note could not locate. The post says the results were "formally verified in Lean" in September. Neither preprint mentions Lean or links a formalization. Hexagon does not accept Lean code in a submission but lets contributors link to formalizations elsewhere (Hexagon, "About"); neither record does. The bootstrap paper's computer assistance is interval arithmetic. The post does not say which results were formalized or where the formalization is.
- Selection. The two results came from 22 candidate problems; the post does not say what happened with the rest.
- Model identity. The post and the preprints name the model differently (see above).
- Scope. The Hawking-Page result is perturbative and planar; the bootstrap result concerns the spinless version of one technique and holds asymptotically in $c$.
Research questions
- Which calculation is right in the disputed diagram class? Recomputing AMMPV's diagrams 3c, 3f and 3n from their printed expressions, as the preprint suggests (Section 8), would settle the one disagreement.
- Does the first-order transition persist at intermediate coupling? A method that follows the transition temperature beyond order $\lambda^2$, such as the Monte Carlo route the preprint mentions (Section 10), would test the post's "most natural possibility."
- What was formalized in Lean? Publishing the formal statements and proofs, with the theorems they cover, would show which parts of the bootstrap result are machine-checked beyond the interval-arithmetic certification.
- How often do such agent runs succeed? Reporting the outcome on all 22 candidate problems, including failures and wrong answers caught, would give a rate rather than two cases.
- Does the conjectured $c^{1/3}$ coefficient hold? Numerical bounds beyond $c = 2048$, or a proof of Conjecture 1, would test the finer structure the paper predicts.
How it connects
- Claude computes a nine-loop amplitude in N=4 super-Yang-Mills (September 2026) is linked from the post as earlier work with the same theory.
- Claude has improved on a longstanding lower bound for the fraction of zeros of the Riemann zeta function (August 2026) and Formalizing Fermat's Last Theorem (September 2026) report AI-found mathematics checked by specialists and in Lean; the comparison is this note's.
- Long-running Claude for scientific computing and Vibe physics: The AI grad student (March 2026) describe earlier agent setups for physics; the connection is this note's.
- System Card: Claude Fable 5.2 & Claude Mythos 5.2 documents the model the post names.
References
- Anthropic. (2026, November 17). Claude solves two open problems in the physics of black holes (page title: "Claude Fable 5.2 solves two open problems in the physics of black holes"). https://www.anthropic.com/research/claude-black-hole-physics record
- Claude (Anthropic). (2026, November 7). The Hawking-Page transition at weak coupling. Hexagon preprint hexagon:2611.00175v1. https://hexagonmath.org/2611.00175v1 record
- Claude (Anthropic). (2026, November 7). The spinless modular bootstrap at large central charge. Hexagon preprint hexagon:2611.00174v1. https://hexagonmath.org/2611.00174v1 record
- Hexagon Mathematics Foundation. About Hexagon (read 21 November 2026). https://hexagonmath.org/about record
- Aharony, O., Marsano, J., Minwalla, S., Papadodimas, K., and Van Raamsdonk, M. (2004). The Hagedorn/deconfinement phase transition in weakly coupled large N gauge theories. Adv. Theor. Math. Phys. 8, 603. arXiv:hep-th/0310285. Cited in the post and the preprint; not read for this note. record
- Hartman, T., Mazáč, D., and Rastelli, L. (2019). Sphere packing and quantum gravity. JHEP 12, 048. arXiv:1905.01319. Cited in the post and the preprint; not read for this note. record
- OpenAI. (2026). Ten Advances in Mathematics and Theoretical Computer Science, Chapter 1. https://cdn.openai.com/pdf/ten-proofs-oai.pdf. Cited in the post and the preprint; not read for this note. record