Anthropic AI Solves Decades-Old Physics Calculations Through Autonomous Agents
San Francisco, Thursday, 19 November 2026.
Running autonomously for six weeks, Anthropic’s Claude solved a 23-year-old black hole calculation, marking a major commercial milestone in shifting AI applications toward automated scientific discovery.
Autonomous AI Agents Tackle Theoretical Physics
Anthropic’s autonomous Claude Fable 5.2 agents, deployed in August 2026, spent six weeks [alert! ‘duration based on multiple sources’] running in the background to address 22 frontier theoretical physics problems, including two longstanding open questions in black hole physics [1][2][8]. The experiment, initiated as a side project by Anthropic researcher Hongbin Chen, marked a shift from traditional AI applications in predictive language processing to autonomous scientific discovery [2][8]. By November 2026, the agents had solved two of the 22 problems, demonstrating collaborative problem-solving capabilities where copies of the same model could read each other’s progress posts and work together on related issues [2][8].
Solving a 23-Year-Old Black Hole Calculation
The first breakthrough involved completing a complex calculation first formulated in 2003, which itself built on a 1983 discovery by Stephen Hawking and Don Page [1][2][8]. Hawking and Page had demonstrated that the transition from a gas of particles to a black hole in a model universe with a negative cosmological constant is abrupt, akin to water boiling at 100°C [1][2][8]. Claude Fable 5.2 not only replicated this result but extended it to prove that the phase transition remains abrupt even when spacetime lacks a definite shape—a scenario where quantum effects dominate [1][2][8]. The calculation, which had remained unresolved for 23 years, was formally verified in Lean in September 2026 and checked by a physicist, adding rigor to the AI’s findings [2][8].
Deriving an Exact Ceiling for Black Hole Mass
The second solved problem involved the modular bootstrap technique, a method used to test candidate theories of quantum gravity [1][2][8]. Previous work by Gerard ‘t Hooft and Leonard Susskind in the early 1990s, later formalized by Juan Maldacena in 1997, had proposed the holographic principle, which states that information within a region of space is proportional to its boundary surface area [1][GPT]. Simeon Hellerman introduced the modular bootstrap in 2009, proving that the lightest black hole in 3D quantum gravity cannot exceed a mass of c/6, where c is the central charge [1][2]. Subsequent analytical work by Tom Hartman, Dalimil Mazáč, and Leonardo Rastelli refined this ceiling to c/8.5 in 2019 [1][2]. Claude Fable 5.2, however, derived an exact ceiling of c/Λ, where Λ ≈ 9.1266, and demonstrated that this is the optimal bound for the simplest version of the modular bootstrap [1][2][8]. This result was verified in Lean and checked by a physicist, confirming its accuracy [2][8].
Implications for AI and Scientific Research
The successful resolution of these two problems underscores a pivotal inflection point in AI capabilities, transitioning from predictive language models to autonomous agents capable of contributing to fundamental scientific research [1][2][8]. Anthropic’s approach involved running a small research group of agent copies, all working collaboratively without special setup, for approximately six weeks [2][8]. The agents not only solved the problems but also wrote the papers with minimal guidance, with a physicist verifying every main result [2][8]. While neither result directly applies to our universe or has immediate practical applications, they represent a significant step toward using AI in high-level R&D automation across technical industries [2][8]. The preprints detailing the findings were published on 16 November 2026 at Hexagon Math, providing open access to the research for further scrutiny and collaboration [1][2][8].
Sources
- www.anthropic.com
- fourweekmba.com
- www.facebook.com
- x.com
- cellcog.ai
- explainx.ai
- www.reddit.com
- www.instagram.com