Not every black hole discovery starts with a telescope. Anthropic says its Claude AI system solved two longstanding theoretical physics problems, including a calculation that researchers outlined in 2003 but never completed.
Announced Nov. 17, the results came from Claude Fable 5.2 agents studying quantum gravity, the effort to reconcile relativity with quantum mechanics. One finding concerns a hypothetical universe shifting from gas to a black hole; the other narrows the search for mathematically consistent theories of gravity.
Neither directly describes black holes in our universe, but both test AI's ability to contribute to complex research.
The first problem: Does a gas become a black hole abruptly?
In 1983, Stephen Hawking and Don Page showed that an idealized universe switches abruptly from a gas of particles to a black hole when heated, much like water boiling.
A related quantum theory showed a similar transition, but researchers could not tell whether it stayed sudden when particles interacted weakly. A calculation proposed in 2003 remained unfinished.
In The Hawking-Page transition at weak coupling, Claude reported completing the calculation and finding an abrupt transition. Intermediate interaction strengths remain unresolved.
The second problem: How heavy could a model black hole be?
The modular bootstrap rules out mathematically inconsistent theories of quantum gravity. In a simplified three-dimensional universe, it sets a ceiling on the mass of the lightest black hole.
An earlier proof set the ceiling at c/8.5, where c roughly counts the model's microscopic building blocks. Numerical estimates suggested a stricter limit.
In The spinless modular bootstrap at large central charge, Claude derived an exact divisor of approximately 9.1266 and showed that this particular technique cannot improve on it. A larger divisor means a lower ceiling. It excludes more theories without identifying the correct one.
A separate AI-generated OpenAI geometry proof also prompted outside mathematical review.
How Claude worked through the problems
Anthropic researcher Hongbin Chen asked Claude to propose open questions. It returned 22, and agents spent about six weeks investigating them.
For the first problem, agents repeated most of the calculations, while a physicist using Claude obtained the same answer in a different way. For the second, Anthropic reported multiple agents converging, formal Lean verification, and physicist review.
The Nov. 7 preprints describe their text as primarily AI-generated. Neither has completed independent peer review. Unlike Claude-designed protein binders tested in laboratories, these results depend on mathematical scrutiny.
What eWeek found: One calculation completed, one mathematical limit established
Comparing the two papers shows that Anthropic's use of the word "solved" refers to different achievements: a completed calculation in one theoretical model and a proven mathematical limit for one analytical method.
Problem | What Claude reportedly established | Where the result stops |
| Hawking-Page transition | An abrupt phase change at weak coupling | Intermediate coupling and other models |
| Modular bootstrap | An exact c/9.1266 ceiling for the simplest spinless method | Other techniques and the correct theory of gravity |
Anthropic acknowledged in its research announcement: "Neither result describes our own universe directly, and neither has practical applications."
What comes next for Claude’s black hole findings
For research teams and organizations exploring AI agents, the significance extends beyond black hole physics.
Anthropic's approach illustrates how multiple agents can investigate difficult problems, check calculations, and work alongside human experts. Similar approaches could be useful in other research-intensive fields, provided the results can be independently evaluated.
However, mathematical progress does not automatically translate into real-world reliability. Unlike Claude's quantum-computer laser work, which engineers tested on hardware, these findings still require broader scrutiny from physicists.
For organizations considering AI agents for complex technical work, the lesson is to look beyond whether an AI produces a convincing answer. Teams also need to understand its assumptions, verify its conclusions, and determine where human expertise remains necessary.
Thinking about putting AI agents to work on your team? Explore The Neuron Academy and its How to Talk to Your AI course to build practical AI skills and learn when human oversight is still needed.


