Claude Fable 5.2 solves two black hole physics problems
Anthropic's AI model resolved long-standing theoretical physics questions about black hole phase transitions and quantum gravity limits using agentic workflows.
Anthropic’s Claude Fable 5.2 has solved two open problems in theoretical physics concerning black holes and quantum gravity. The breakthrough, published in preprints on November 17, 2026, demonstrates the capability of agentic AI workflows to handle complex mathematical derivations that have stumped human researchers for decades.
What happened
The first problem involves a calculation set up by physicists more than twenty years ago but never completed. It concerns the behavior of theoretical universes when heated. In 1983, Stephen Hawking and Don Page showed that such a universe undergoes an abrupt phase transition from a gas of particles to a black hole at a specific temperature. However, their calculation relied on general relativity, which breaks down at small scales where spacetime becomes blurry. In 1999, researchers found a similar transition on the quantum side of the holographic duality, but it was unknown whether this transition was also abrupt or smooth. Claude completed the missing calculation, proving that the phase transition remains abrupt even when spacetime loses its definite shape.
The second problem relates to the modular bootstrap method, a technique used to test candidate theories of quantum gravity by ruling out inconsistent ones. Since 2009, physicists have used this method to set an upper limit on the mass of the lightest black hole in simplified two-dimensional universes. While numerical calculations hinted at a precise ceiling, no one could prove it mathematically. Claude Fable 5.2 identified a pattern in existing numerical results and borrowed a method from sphere-packing mathematics to calculate the exact limit. It proved that this ceiling cannot be lowered, providing a rigorous boundary for future theories.
These results were achieved through a background experiment started in August by Anthropic researcher Hongbin Chen. He tasked Claude with listing open physics problems, then deployed separate agents to work on them. The agents collaborated, checked each other’s work, and ran for about six weeks without special setup. The findings were formally verified in Lean, a computerized proof checker, in September before being written up with minimal human guidance.
How it works
The solution relies on holographic duality, a principle suggesting that a universe with gravity can be described by a quantum theory on its boundary, much like a hologram encodes 3D information on a 2D surface. This allows physicists to translate hard questions about gravity into potentially easier questions about quantum particles. Claude operated as a small research group of copies, working independently on different parts of the calculations and cross-referencing known results.
For the bootstrap problem, the AI recognized a mathematical pattern that human researchers had missed. It applied a technique from high-dimensional sphere packing, a field where another AI model had recently made progress. By adapting this method, Claude derived the exact limit for the black hole mass ceiling. The system’s ability to connect disparate areas of mathematics and physics enabled it to bypass the computational barriers that had previously stalled progress.
Key details
- Claude Fable 5.2 solved a 20-year-old incomplete calculation regarding the Hawking-Page phase transition in black hole formation.
- The AI proved that the transition from particle gas to black hole is abrupt, even in regimes where spacetime is not well-defined.
- It determined the exact mathematical limit for the lightest black hole mass in simplified two-dimensional universes using the modular bootstrap method.
- The research was conducted by autonomous agents running in the background for six weeks, starting in August 2026.
- Results were formally verified using the Lean proof assistant in September 2026.
- The work confirms that certain black hole features are built into the underlying quantum theory, not just artifacts of general relativity.
Why it matters
For software engineers and AI developers, this milestone highlights the growing proficiency of agentic workflows in handling deep, multi-step reasoning tasks. Unlike previous AI applications that focused on code generation or data summarization, this achievement involves novel mathematical discovery and formal verification. It suggests that future AI systems may assist in scientific research by identifying patterns across disparate fields and executing complex logical chains that exceed human working memory limits.
The use of formal verification tools like Lean ensures that the AI’s outputs are mathematically rigorous, addressing concerns about hallucinations in critical domains. This combination of generative AI for hypothesis generation and symbolic AI for verification represents a robust architecture for high-stakes problem solving. While these specific physics results have no direct practical application, the methodology demonstrates how AI can accelerate theoretical exploration in fields where data is scarce but logical constraints are tight.
What you can do
- Study the structure of agentic workflows to understand how independent agents can collaborate on complex tasks.
- Explore formal verification tools like Lean to see how mathematical proofs can be automated and checked.
- Review the preprints associated with the findings to understand the specific mathematical techniques employed.
- Consider how pattern recognition across different domains could be applied to optimization problems in your own software projects.
- Monitor developments in AI-assisted scientific discovery to identify opportunities for integrating similar workflows in technical research.
- Evaluate the reliability of AI-generated logic by incorporating automated testing and verification steps into your development pipeline.



