Anthropic says Claude agents solved two open black hole problems
Anthropic reported Claude Fable 5.2 closed two unsolved theoretical physics problems on black holes.

· Last verified: 19 Nov 2026 (Anthropic)

Summary
- Anthropic reported Claude Fable 5.2 closed two unsolved theoretical physics problems on black holes.
- One finished a 2003 calculation left incomplete; the other tightened a 2011-era mathematical bound.
- A physicist reviewed every main result and Lean verified the bootstrap proof.
The latest
Two open problems in black hole theory were solved by Claude agents, Anthropic's science team said in a November 17, 2026 post, with results checked by a proof assistant and reviewed by a physicist. One agent completed a quantum gravity calculation abandoned since 2003; another derived a sharper limit on the lightest possible black hole mass in three-dimensional gravity.
Details
- The first problem: Anthropic said Claude settled whether the Hawking-Page transition in N=4 supersymmetric Yang-Mills theory is abrupt or smooth at weak coupling. The calculation required a three-loop coefficient and thousands of diagrams. Claude found the transition is abrupt, matching Hawking and Page's 1983 result.
- The method: A small set of Claude instances ran the Hawking-Page work, with nearly every step computed twice independently, according to Anthropic. A physicist running Claude separately on the same problem reached a matching result; a colleague noted the Anthropic agent got there first.
- The second problem: On the modular bootstrap for three-dimensional gravity, Anthropic said Claude derived an exact ceiling on the lightest black hole mass, M₁ ≤ c/Λ with Λ approximately 9.1266, and showed the bound is optimal for that form of the method.
- The comparison: Earlier numerical work pointed to roughly c/9.1, while the best previously proven bound was c/8.5, credited to Hartman, Mathys and Rastelli. The new ceiling still falls short of c/12, the threshold where black holes begin to appear, Anthropic said.
- The failures: Anthropic disclosed that 61 parallel Claude attempts failed on the bootstrap problem before a separate agent connected the method to a sphere-packing result recently proven by an OpenAI model. Two further agents independently arrived at the same answer.
- The origin: Anthropic researcher Hongbin Chen tested Claude on the work as a side project in August, the company said. Claude proposed 22 candidate problems, and separate agents worked on them for about six weeks.
- The verification: Lean, a computer proof checker, verified the modular bootstrap results in September, according to Anthropic. Claude wrote the paper with limited human direction, and a physicist reviewed every major result before publication.
- The limits: Anthropic stated plainly that neither result describes our universe directly and that neither has practical applications. The two papers were posted as preprints, numbered 2611.00175v1 for Hawking-Page and 2611.00174v1 for the modular bootstrap.
- Also: Neither result describes our universe or carries practical applications, Anthropic said.
Background
The Hawking-Page transition, identified in 1983, describes the point at which a gas of particles collapses into a black hole. The 2003 weak-coupling calculation that would test it at that regime was never completed until now, according to Anthropic.
Between the lines
The disclosed failure count is the most revealing number here: 61 parallel attempts collapsed before a different agent found the sphere-packing link. That suggests the breakthrough came from routing between methods rather than raw computation. Anthropic's own framing — no practical applications, no description of our universe — positions the claim as a capability demonstration, not a physics event.
What's next
Independent physicists have not yet publicly assessed the two preprints. Watch for peer response to the Λ ≈ 9.1266 bound, and whether other labs replicate the agent-driven workflow on comparable open problems.
Source: Anthropic