Anthropic Says Claude Fable 5.2 Solved Two Black Hole Physics Problems
Anthropic released research this week indicating that its Claude Fable 5.2 model autonomously resolved two long-standing problems in theoretical physics concerning black holes.
The firm published the findings in a blog post this week, positioning the model’s performance as a demonstration of its scientific reasoning capabilities. The Tech TL;DR:
- Claude Fable 5.2 autonomously solved two physics problems related to black hole formation and quantum gravity that had remained open for decades.
- The AI utilized a “bootstrapping” technique to identify patterns in mathematical calculations that human researchers had previously missed.
- These findings do not currently apply to our universe, as they focus on theoretical “anti-de Sitter” space and simplified quantum gravity models.
The first problem involved the Hawking-Page phase transition, a concept originally described by Stephen Hawking and Don Page in 1983. This theory posits that a hot gas in a specific type of universe—modeled as a giant bowl—will collapse into a black hole at a certain temperature. While Hawking and Page analyzed this in a vast universe governed by general relativity, Anthropic reported that Claude confirmed this phase transition remains abrupt even in much smaller, differently structured theoretical universes. According to the company, this suggests that black hole formation is an inherent feature of underlying quantum theory rather than a byproduct of relativity alone. The second finding concerns the “modular bootstrap” method, which physicists use to determine the minimum possible mass of a black hole in a two-dimensional universe. Mashable reported that researchers have spent years attempting to lower the ceiling for this minimum mass by creating increasingly complex bootstrap functionals.
Anthropic said that Claude Fable 5.2 identified a pattern within these calculations that had eluded human physicists, allowing the model to calculate a precise lower bound that cannot be further reduced. “Claude came back hours later with 22 candidates, and Hongbin set up separate agents to work on each one,” Anthropic said in its research post, detailing how researcher Hongbin Chen initiated the project in August. These agents operated autonomously for approximately six weeks, communicating and iterating on their findings without manual intervention.
The company emphasized that while these results do not solve the overarching mystery of quantum gravity, they validate the use of large language models for complex analytical tasks in academic research.
The methodology relied on the model’s ability to parse existing literature and apply mathematical tools from disparate areas of physics. By running these agents in the background of their standard compute infrastructure, Anthropic demonstrated a workflow that mirrors modern continuous integration practices, where autonomous processes handle iterative problem-solving.
# Example of the bootstrap logic structure
def calculate_bootstrap_ceiling(theory_params):
# Analyze the functional space for potential lower bounds
# Claude Fable 5.2 identified pattern: localized variance in 2D manifolds
if pattern_detected(theory_params):
return precise_lower_bound
return estimate_ceiling
For now, the research serves as a proof-of-concept for AI-driven scientific discovery, though it stops short of providing a unified theory of quantum gravity. Further progress will depend on whether these models can successfully transition from theoretical “toy universes” to the complex, non-linear variables present in our own physical reality.
Anthropic, a public benefit corporation founded in January 2021 by former OpenAI researchers, developed the Claude model family used in this research. The company was valued at $965 billion in a May 2026 capital raise and went public that November. Its researchers include Nobel laureate John M. Jumper and former OpenAI safety leads Jan Leike and John Schulman.