Anthropic · 2026-11-17 · major
Claude solves two open black hole physics problems — papers by Claude agents
In short
Anthropic says Claude agents solved two open problems in black hole physics: the weak-coupling Hawking-Page transition and the exact modular bootstrap ceiling, about c/9.1266. Both papers were written by Claude and checked by a physicist.
Anthropic · 2026-11-17 · 2 sources checked

Claude agents finished a 2003 black hole calculation and found an exact bound physicists had only estimated.
Key specs
| Bootstrap ceiling | c / 9.1266 |
|---|
Quick facts
| Project lead | Hongbin Chen, Anthropic |
|---|---|
| Problem 1 | Hawking-Page transition in N=4 super Yang-Mills at weak coupling |
| Problem 2 | Optimal modular bootstrap bound for 3D gravity |
| Previous best bound | c / 8.5 proven; numerics near c / 9.1 |
| Run time | About six weeks of background agents on 22 candidate problems |
| Checking | Physicist review; bootstrap result verified in Lean |
| Papers | Hexagon 2611.00174 and 2611.00175, CC BY 4.0 |
What is it?
Anthropic's research post reports two results in theoretical physics produced by Claude. First, Claude completed a calculation set up in 2003 but never finished, showing that the Hawking-Page phase transition in N=4 super Yang-Mills stays abrupt (first order) at weak coupling. Second, it found the exact ceiling on the lightest black hole's mass in the spinless modular bootstrap, about c/9.1266, and proved it is optimal for that method.
How does it work?
Researcher Hongbin Chen asked Claude to list open problems in black hole physics and the bootstrap method; it produced 22, and separate agents worked on them in the background for about six weeks. For the phase transition, a group of Claude copies did nearly every piece of the calculation twice, independently. For the bound, Claude spotted a pattern in numerical results and used a spectral curve from another area of mathematics to get the closed form, after 61 earlier parallel attempts had failed.
Why does it matter?
Physicists had a proven bound of c/8.5 and numerical estimates near c/9.1; the new result pins the exact value. The papers name Anthropic as contributor and say Claude wrote them under light guidance, with a physicist checking the main results, so this is AI-led research rather than AI-assisted. The method still cannot reach c/12, where black holes first appear, so the bigger question stays open.
Who is it for?
physicists, mathematicians and people tracking AI for science
Frequently asked questions
- Which Claude model solved the black hole problems?
- The modular bootstrap paper says an internal Claude research model first obtained the main result and that Claude Opus 5.5 later reproduced it. The Hawking-Page paper credits an internal Claude research model prompted by a physicist. Anthropic's post adds that a separate agent reached the same phase-transition answer before a physicist who was also using Claude.
- How were Claude's black hole physics results checked?
- Anthropic says a physicist checked the main results. In the phase-transition work, independent Claude agents did nearly every piece of the calculation twice, and the paper reproduces known results such as the two-loop Hagedorn temperature. The bootstrap result was reached independently by two other agents and formally verified in Lean in September.
- Does Claude's modular bootstrap result prove pure 3D gravity exists?
- No. Claude's exact ceiling of about c/9.1266 is the best this version of the modular bootstrap can do, and Anthropic says the method cannot reach c/12, where black holes first appear. Settling that question needs other ingredients. The result does replace the earlier proven bound of c/8.5 by Hartman, Mazáč and Rastelli.
- Where can I read Claude's black hole physics papers?
- Both papers are posted on Hexagon under a CC BY 4.0 license: 'The spinless modular bootstrap at large central charge' (2611.00174) and 'The Hawking-Page transition at weak coupling' (2611.00175). Version 1 of each was posted on November 7, 2026, ten days before Anthropic's announcement, and the phase-transition paper is labeled as primarily AI-generated text.
