Anthropic has published how its model solved two open problems in theoretical physics, both about black holes. In a research post dated 17 November, it says an Anthropic researcher, Hongbin Chen, asked Claude for open problems, got “22 candidates”, and set up a separate agent for each one. The agents ran “for about six weeks”.
Two of the 22 problems were solved, about 9% of the list by our arithmetic. Anthropic first announced the two results on 9 November; this post explains the method behind them. It is Anthropic’s own account, published in the week of its IPO, and we have not verified the physics.
Business Pill · The Model and the Harness
A short Business Engineer explainer on the difference between an AI model and the harness around it. It teaches the general idea only and does not discuss this research.
The key insight: Anthropic says 22 Claude agents worked on open physics problems for about six weeks and solved two, about 9% by our arithmetic, with one result formally verified in Lean, per its 17 November post.
How the Agents Worked
The agents, the post says, “could read each other’s progress posts and collaborate when working on related problems” and ran “in the background, with no special setup”. Chen started the work in August as a side project, according to Anthropic.
On the first problem, Claude “ran a small research group of agents, all copies of itself”, and “almost every piece of the calculation was done twice, by agents acting independently”, the post says.

Problem One: the Hawking-Page Transition
In 1983, Stephen Hawking and Don Page showed that a hot gas in a model universe turns abruptly into a black hole. A 2003 group set out a calculation to test whether that jump also holds where spacetime has no definite shape, but it was never completed. Claude’s answer, per the post: the transition is abrupt there too.
Anthropic adds that a physicist who had independently used Claude on the same problem later shared his result, and that “the two solutions used different methods, so they were a useful check on each other”.
Problem Two: the Modular Bootstrap
The second result concerns a ceiling on how heavy the lightest black hole can be in a simplified universe, written as M1 ≤ c/Λ. Earlier numerical work suggested a ceiling near c/9.1, and the best analytical proof stood at c/8.5, the post says. Claude derived an exact value: Λ is “approximately 9.1266”.
Getting there took several agents. One spotted a pattern in published numerical results; another, working on a sphere-packing problem, turned it into a method that reproduced a result recently proved by an OpenAI model. An earlier Claude agent had “tried, and failed, to solve in 61 parallel attempts”, according to the post.
How the Results Were Checked
For the bootstrap result, Anthropic says that “in September, the results were formally verified in Lean, a computerized proof checker”, and that Claude “wrote the paper with minimal guidance, and a physicist checked every main result”. For the Hawking-Page result, the checks it describes are the agents’ double computation, comparison with earlier literature and the outside physicist’s matching answer. The two preprints are linked from the post (one, two). Our 9 November report on the launch claims is here.
The post is careful about scope: “Neither result describes our own universe directly, and neither has practical applications”. For other AI-for-research results this week, see Google’s agent-harness paper and Epoch’s staff-bar test.
The Structural Read
As we read the post, the method is the news: one researcher, a list the model drew up itself, and agents that shared progress without a special setup.
As we read it, the 20 problems without a result matter as much as the two solved, and the post does not describe them.
Anthropic research post, 17 Nov 2026
“Neither result describes our own universe directly, and neither has practical applications”
Three Implications
FOR RESEARCH LABS Anthropic describes agents that read each other’s progress posts and collaborated on related problems, per the post.
FOR VERIFICATION For the bootstrap paper, the post says the results were formally verified in Lean and a physicist checked every main result.
FOR COMPETITORS One Claude method reproduced a sphere-packing result recently proved by an OpenAI model, Anthropic says.
What Is Not Established
We have not checked the calculations or the Lean proofs. The post does not say how much compute the six weeks used, what happened to the other 20 problems, or whether the preprints have been peer reviewed. We did not contact Anthropic.
The Bottom Line
Anthropic’s post says 22 Claude agents worked on open physics problems for about six weeks and two problems were solved, with the bootstrap result verified in Lean and checked by a physicist, per the company.
97,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. Quotes and figures are from Anthropic’s 17 November research post, read in full. The 9% share (2 of 22) is our arithmetic. We did not verify the physics, read the full preprints or contact Anthropic.
Sources: Anthropic, Claude Fable 5.2 solves two open problems in the physics of black holes, 17 Nov 2026 · Preprint 2611.00174, linked from the post · Preprint 2611.00175, linked from the post









