Claude solved two open problems in black hole physics
Claude and black hole physics: Anthropic agents finished a calculation left open since 2003 and pinned down a bound on the lightest black hole. How it happened.
This article in other languages:Polski
In this article
In a post titled “Claude solves two open problems in the physics of black holes” (dated November 17, 2026), Anthropic describes how Claude solved two open problems in theoretical physics. Both concern black holes and the search for a quantum theory of gravity. The preprints are public: the first is on the Hawking-Page transition and the second on the so-called “modular bootstrap”.
Claude and black hole physics: what exactly was computed?
The first problem is more than 20 years old. In 1983, Stephen Hawking and Don Page showed that in a model universe (anti-de Sitter space), a hot gas of particles abruptly gives way to a black hole above a certain temperature, much as water starts to boil at 100°C. Their calculation rests on general relativity, so it only works for a large universe.
For a small one you have to compute on the quantum side of the holographic duality, in N=4 super Yang-Mills theory. Since 1999 it has been known that a phase transition exists there too, but not whether it is abrupt or smooth. In 2003, Ofer Aharony and collaborators showed that a single number settles it, one that only appears at three loops and takes many thousands of diagrams to compute. Nobody finished that calculation.
Claude finished it. The paper’s abstract gives the coefficient as b = −1.48 × 10⁻³ λ², and the transition is first-order, meaning abrupt, just as in Hawking and Page. Nearly every number in the computation was obtained in two independent ways.
The second problem is about simplified gravity in three spacetime dimensions. Since 2009, physicists have been working out a “ceiling” there on the mass of the lightest black hole, written as M₁ ≤ c/Λ (c is the central charge). Theories that break through the ceiling are internally inconsistent and get thrown out. The best ceiling proven so far is c/8.5 (Hartman, Mazáč and Rastelli), and numerical work pointed to somewhere around c/9.1.
According to the second preprint, the exact value is Λ = 12π²/(π² + 4 arcsinh² 1) ≈ 9.1266. Claude also proved that the simplest version of this method (the one that ignores black hole spin) cannot go any lower.
How did the agents work?
It was meant to be a side project. In August, Anthropic researcher Hongbin Chen asked Claude for a list of interesting open problems in black hole physics and the bootstrap method. A few hours later he had 22 proposals and assigned a separate agent to each. The agents ran in the background for about six weeks (with no special setup, Anthropic says), could read each other’s progress notes, and collaborated on related topics.
The post describes two solved problems and says nothing about the other 20.
The bootstrap result came about partly by accident. The agent assigned to the problem spotted a pattern in numerical data that physicists had missed. A different agent, one working on sphere packing in high dimensions, turned it into a method, and that method reproduced a result recently proven by an OpenAI model. Half of the final proof follows the same strategy as OpenAI’s. An earlier Claude agent had tried the same problem in 61 parallel attempts and failed.
In September the results were formally verified in Lean (I wrote about proofs like that in the post on the exponential time hypothesis). Claude then wrote the paper, and a physicist checked every main result.
Which Claude model did it?
Here the post and the preprints use different words. The post names Claude Fable 5.2, while the provenance statements in both papers refer to an “internal Claude research model”. The agents started in August and Fable 5.2 launched on November 9, so it was probably a version from before the public release doing the work.
One sentence in the bootstrap preprint matters more: the publicly available Claude Opus 5.5 later reached the same result independently. That result, at least, is within reach of a model we can all use.
What to watch for
First, the status of the papers. These are version 1 preprints from November 7, labeled on Hexagon as text “mostly generated by AI”. For the Hawking-Page transition, though, there are two independent calculations: a physicist who was himself using Claude reached the same result by a different method.
Second, what the results don’t say. Anthropic notes that neither describes our universe directly and neither has practical applications. The c/9.1266 ceiling also falls short of c/12, the mass at which black holes appear at all, and getting there takes different tools.
Third, the difference between the two results. The first is a known calculation, too long and too error-prone to do by hand. The second needed a new idea: noticing a pattern in someone else’s data and reaching for the “spectral curve”, a tool known from other areas of mathematics and physics.
FAQ
Did Claude solve quantum gravity?
No. According to Anthropic, neither result settles the question of quantum gravity or comes close to it. Both concern simplified model universes.
Where can I read Claude’s black hole papers?
On Hexagon: “The Hawking-Page transition at weak coupling” and “The spinless modular bootstrap at large central charge”. Anthropic released both under a CC BY 4.0 license.
Were humans involved in the work?
Yes, but only to a limited extent. An Anthropic researcher launched the agents and picked the topic, and a physicist checked the main results and gave brief guidance while the bootstrap paper was being written. The solutions themselves, Anthropic says, were produced largely autonomously.