The Pulse
Anthropic says Claude solved two black-hole physics problems
Anthropic says Claude Fable 5.2 completed a long-standing calculation about black-hole formation and found a limit for testing theories of quantum gravity. The company published its account alongside two preprints.

Anthropic says its Claude Fable 5.2 model solved two open theoretical physics problems involving black holes, completing a long-standing calculation in one and finding a limit on a method for testing quantum gravity theories in the other. In a research post published November 17, the company described the work and linked to preprints of both results.
Both findings concern theoretical universes rather than the universe people inhabit, and neither has practical applications, the company said. They address questions in the effort to reconcile general relativity, which describes large-scale phenomena, with quantum mechanics, which describes very small ones.
A transition in a smaller universe
In the first problem, physicists wanted to know what happens as a theoretical universe heats up. In 1983, Stephen Hawking and Don Page showed that, in a large enough universe of this kind, a gas changes abruptly into a black hole above a certain temperature.
That calculation relied on general relativity. At a much smaller scale, spacetime no longer has a definite shape, and the same calculation cannot be performed in the same way on the gravity side of the theory.
In 1999, physicists found that a transition also occurs in that smaller-scale regime, but did not establish whether it was abrupt or smooth. A group of physicists set out a calculation to answer that question in 2003, but it was not completed.
Claude completed the calculation and found that the transition is abrupt there too, Anthropic said. The model worked through almost every part independently using copies of itself and checked its mathematics against known results, according to the company.
A limit on testing theories
The second problem involves a simplified theoretical universe with two dimensions of space rather than three. Physicists use a method called bootstrapping to rule out candidate theories of quantum gravity by showing that they are internally inconsistent.
Since 2009, researchers have used a version of the method to set an upper limit on the mass of the lightest black hole in such a universe. A theory that exceeds the limit can be ruled out. Numerical calculations had suggested where the best possible limit lay, but had not proved it.
According to the company, Claude identified a pattern in those numerical results and used a method from elsewhere in mathematics and physics to calculate and prove the precise limit for the simplest version of the technique.
How the agents worked
In August, Anthropic researcher Hongbin Chen asked Claude to list open problems in theoretical physics, focusing on black holes and the bootstrap method. The model returned 22 candidates, the company said.
Chen assigned separate agents to the problems; they could read one another’s progress posts and collaborate on related work. The agents ran in the background for about six weeks, according to Anthropic.
Weeks after one agent reached a solution to the transition problem, a physicist who had independently used Claude on it shared the same answer, the company said. The two solutions used different methods.
For the bootstrap result, one agent spotted the numerical pattern and another, working on a related sphere-packing problem, developed it into a method. An earlier Claude agent had failed to solve the bootstrap problem in 61 parallel attempts, Anthropic said; two other agents independently reached the same answer as the later effort.
In September, the bootstrap results were formally verified in Lean, a computerized proof checker, according to the company. A physicist checked every main result after Claude wrote the paper with minimal guidance.
Source
Companies in this story
Related stories
- Anthropic resolves Claude errors after API and Code disruption
- Anthropic advises deleting "think hard" lines from Claude Opus 5.5 prompts
- Anthropic opens self-hosted Claude Managed Agents to private beta
- Anthropic adds beta sheets and whiteboards to paid Claude plans
- Anthropic publishes recommendations for designing, running and reporting AI safety tests