Anthropic has announced that Claude Fable 5.2 has independently solved two significant open problems in theoretical physics, specifically targeting the intersection of general relativity and quantum mechanics. The findings, detailed in a research post published on November 19, 2026, involve completing a calculation on black hole phase transitions that had remained unsolved for over two decades, and determining the exact limit of a technique used to test candidate theories of quantum gravity. This marks a pivotal moment where large language models are not merely assisting in scientific discourse but are actively contributing novel mathematical proofs to foundational physics.

The Hawking-Page Transition Puzzle

The first breakthrough concerns the Hawking-Page phase transition, a concept introduced by Stephen Hawking and Don Page in 1983. They demonstrated that in anti-de Sitter spacetime, a universe filled with a gas of particles abruptly transitions into one containing a black hole when heated past a critical temperature. While this transition is clear in the gravity description, it was unknown whether this abruptness persists in the dual quantum description, specifically within N=4 super Yang-Mills theory at weak coupling. In 2003, physicists outlined the necessary calculation to settle this, but it involved thousands of complex diagrams and was never completed by human teams. Claude Fable 5.2 solved it by deploying a swarm of self-collaborating agents that independently computed the vast majority of the calculation's components, ultimately proving that the phase transition is indeed abrupt in the quantum regime as well.

Proving the Modular Bootstrap Limit

The second achievement involves the modular bootstrap, a method introduced by Simeon Hellerman in 2009 to rule out inconsistent theories of quantum gravity in three-dimensional spacetime. Physicists have long sought to lower the ceiling on the mass of the lightest black hole in these model universes, a task that narrows the search for valid quantum gravity theories. While numerical results hinted at a precise limit, they lacked formal proof. Claude Fable 5.2 identified a pattern in these numerical results that human researchers had missed and borrowed a mathematical method from sphere-packing problems in high-dimensional geometry. This allowed the model to calculate the exact ceiling for the simplest version of the bootstrap technique and prove that it could not be lowered, a result that cross-references a similar proof recently achieved by an OpenAI model.

Autonomous Research Methodology

The methodology behind these discoveries reveals a shift in how AI agents conduct scientific research. Anthropic researcher Hongbin Chen initially prompted Claude to list open problems in theoretical physics, receiving 22 candidates. He then deployed separate agents to work on each problem for approximately six weeks in the background. These agents were not isolated; they read each other’s progress and collaborated on related tasks. For the Hawking-Page problem, the solution was verified independently by a human physicist who used Claude separately, with the two distinct methods serving as cross-checks. The final proofs were formally verified in Lean, a computerized proof checker, before Claude drafted the paper with minimal guidance.

Key Takeaways

  • Claude Fable 5.2 completed a 20-year-old calculation on the abruptness of the Hawking-Page phase transition in weak-coupling quantum theories.
  • The model proved the exact limit of the modular bootstrap for 3D gravity by identifying a pattern missed by human physicists.
  • The research was conducted by a swarm of autonomous agents that collaborated and cross-verified results over a six-week period.
  • The findings were formally verified using the Lean proof assistant, ensuring mathematical rigor independent of human intuition.

The Bottom Line

This isn't just a benchmark win; it's a signal that LLMs are moving from pattern matching to genuine mathematical discovery, potentially outpacing human researchers in tedious, high-complexity theoretical proofs.