News 4 min read

Claude Fable 5.2 solves two open problems in black hole physics


Anthropic says Claude Fable 5.2 has solved two open problems in theoretical physics, both about black holes and quantum gravity. In a research post published on November 17, 2026, the company describes a group of Claude agents that ran in the background for about six weeks and produced two preprints: one completes a calculation physicists set up in 2003 and never finished, the other finds and proves the exact limit of the simplest version of a technique used to rule out theories of 3D quantum gravity. For developers, it is one of the clearest public examples so far of long-running, multi-agent work on a hard technical problem, and of how the results were checked.

What’s new

Problem 1: the Hawking-Page transition at weak coupling. In 1983 Stephen Hawking and Don Page showed that a hot universe of the kind used in these models switches abruptly from a gas of particles to a black hole, a phase transition like water boiling. Their calculation only works where general relativity holds. On the quantum side of holographic duality, in N=4 super Yang-Mills theory, physicists found in 1999 that a transition also happens when particles barely interact (weak coupling), but not whether it is abrupt or smooth. In 2003, Ofer Aharony and colleagues set out the calculation that would decide it; the answer hinged on one number that appears three loops in and takes many thousands of diagrams to compute. It was never completed for N=4 super Yang-Mills.

Claude completed it and found the transition is abrupt. Anthropic says the transition is now known to be abrupt at both weak and strong coupling, and calls it “the most natural possibility” that it stays abrupt in between, which is not proven.

Problem 2: the modular bootstrap ceiling. The modular bootstrap, introduced by Simeon Hellerman in 2009, rules out inconsistent theories of 3D gravity by putting a ceiling on the mass of the lightest black hole, written M₁ ≤ c/Λ, where c is the central charge. Hellerman proved c/6. Numerical work suggested the simplest version of the method could reach about c/9.1, while the best proof, by Tom Hartman, Dalimil Mazáč and Leonardo Rastelli, was c/8.5.

Claude spotted a pattern in how the numerical ceiling shifted as c grew, brought in a tool called a spectral curve that had not been applied to this problem, and proved a ceiling with Λ given by an exact formula, approximately 9.1266. It also proved this is the lowest this version of the method can go, and therefore that this version of the method cannot, on its own, reach c/12, where black holes first appear.

The papers are on hexagonmath.org: the Hawking-Page preprint and the modular bootstrap preprint. Anthropic does not mention peer review, and says neither result describes our own universe directly or has practical applications.

How the agents worked

The setup is the part most relevant to people who build with AI agents. According to Anthropic:

  • In August, Anthropic researcher Hongbin Chen, as a side project, asked Claude to list open problems in theoretical physics focused on black holes and the bootstrap. It came back with 22 candidates.
  • Chen gave each problem its own agent. The agents could read each other’s progress posts and collaborate on related problems, and ran “with no special setup” for about six weeks.
  • For the Hawking-Page problem, almost every piece of the calculation was done twice by independent agents and checked against known results. Weeks after an agent reached its answer, a physicist who had separately used Claude on the same problem shared a matching result found by a different method.
  • For the bootstrap, an agent working on a related sphere-packing problem turned the spotted pattern into a method, which reproduced a sphere-packing result recently proved by an OpenAI model. An earlier Claude agent had failed on the bootstrap problem in 61 parallel attempts; two other agents later reached the same answer independently.
  • The bootstrap results were formally verified in Lean, a proof checker, in September. Claude then wrote the paper “with minimal guidance”, and a physicist checked every main result.

What it means for developers

  • Verification did the heavy lifting. The claims rest on duplicated independent work, comparison with published results, a Lean proof and a human expert check, not on a single model run. That is a pattern worth copying for any agent output you need to trust.
  • Shared progress notes. Agents that can read each other’s notes found a method in one problem and reused it in another, which is the case for giving parallel agents a shared workspace.
  • No new product. The post announces research, not a feature: there is no new model, API or pricing. Claude Fable 5.2 itself is the model Anthropic released on November 9, available to all users. For how Anthropic’s models are speeding up its own work, see our post on the METR report.

How it compares

Anthropic’s post itself points to OpenAI’s recent sphere-packing proofs, and half of Claude’s bootstrap proof (that the ceiling can’t be lowered) follows the strategy of that OpenAI proof. Anthropic also links Claude’s earlier work on N=4 super Yang-Mills, calculating particle interactions in that theory to a new level of precision. Anthropic’s own summary: “Neither result settles the question of quantum gravity, or even comes close to doing so, but each pushes physics forward in a longstanding area of research.”

Sources