AI NewsModels & agentsAnnouncement
Anthropic says Claude solved two open problems in black hole physics
Anthropic says it handed Claude Fable 5.2 twenty-two open problems in black hole physics. Agents ran for about six weeks and closed two of them.

Image: Anthropic
Why it mattersThe work is a test of how far an unsupervised agent run can go on a hard research task. The method matters more than the finding for teams building long agents.
A small group of Claude agents closed two open questions in black hole physics. They ran on their own for about six weeks. One was a calculation physicists set up in 2003 and never finished. The other was a bound on the lightest black hole in 3D quantum gravity. That bound had sat between a numerical guess and a weaker proof for years.
The post is on Anthropic's own research site. The work is now a paper on the pre-print site hexagonmath.org. Anthropic researcher Hongbin Chen gave the model twenty-two open problems in August 2026. He let the agents work. Claude Fable 5.2 was the model. A physicist checked every main result by hand before the post went out. Both proofs were also run through Lean in September. Lean is a program that checks a proof line by line.
What Claude did on the first problem
The first problem was the Hawking-Page transition at a quantum regime. A 2005 paper closed a smaller version of the same calculation. The full version was more than ten times as large. It had been open since 2003. Claude ran it as a group. Agents did each piece twice and checked the answer against earlier work. The result holds at both the weak and the strong edge of the theory. The transition from no black hole to a black hole is abrupt. That is what Hawking and Page guessed in 1983.
What Claude did on the second problem
The second problem was a ceiling on the lightest black hole in 3D quantum gravity. Numerical work had put the ceiling near c over 9.1. The best written proof had reached only c over 8.5. Claude saw a pattern in the numbers that people had missed. It brought in a tool called spectral curves. Nobody had used that tool on this problem. The number Claude found is c over 9.1266. Claude also proved this is the lowest a proof of this shape can go. One agent spotted the pattern. Another, working on a different problem, turned it into a method. Two more checked the final answer on their own.
What the method actually was
Anthropic does not claim the model worked alone. The researcher chose the problems. He asked follow-up questions and read the drafts. The agents did the calculations and wrote the notes. Each main step was done twice. Each was then compared with the field's existing work. Every main result had a physicist check it. For the modular bootstrap result, a Lean proof was built in September.
Anthropic frames this as evidence that a current model can help on a frontier research task. The setup has to be right. An expert in the loop. Separate agents checking each other. A formal proof tool at the end. The post says the agents ran for weeks on twenty-two problems and closed two of them. Most of the runs did not close anything.
For a team building long agents, the interesting part is the setup. Two agents check the same answer. A verifier runs at the end. Here that is Lean. In software it would be a test suite. An expert reads the output before it leaves the lab. That shape is one another team can copy on a different task.
Source
Claude solves two open problems in the physics of black holes, Anthropic.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.


