Key takeaways
- Anthropic says Claude agents, running for about six weeks, showed the Hawking-Page transition in N=4 super Yang-Mills stays abrupt at weak coupling, finishing a calculation set up in 2003.
- A second paper says the spinless modular bootstrap limit is exactly 12π²/(π² + 4 arcsinh²1), about 9.1266, and that this version of the method cannot do better.
- Both are preprints on Hexagon, checked by Anthropic, Claude and unnamed physicists, with no peer review or published outside check yet.
Anthropic says Claude has solved two open problems in black hole physics. In a research post dated November 17, 2026, the company describes how copies of the model, running as parallel agents for about six weeks, finished a calculation physicists set up in 2003 and found the exact limit of a 15-year-old technique for testing theories of quantum gravity. Both results are written up as preprints, not yet peer reviewed, and neither says anything directly about our own universe.
The background: black holes as a test for quantum gravity
General relativity describes gravity and big things like stars. Quantum mechanics describes small things like atoms. They don't fit together, and a theory that joins them, quantum gravity, is one of physics' biggest unsolved goals. Black holes are a test case because both theories matter there.
Both problems use holographic duality: some model universes with gravity can be described exactly by a quantum theory with no gravity, living on the universe's boundary. That turns hard gravity questions into quantum ones. These model universes are anti-de Sitter spaces, where everything gets pulled back toward the middle, like a ball inside a bowl. Ours is not like that.
Problem one: does a black hole form abruptly in a tiny universe
In 1983, Stephen Hawking and Don Page showed that a heated bowl-shaped universe switches suddenly from hot gas to a black hole, the way water boils. Their calculation only works when the universe is large enough for relativity to hold.
When the universe is so small that space has no definite shape, only the quantum side works. Here that is N=4 super Yang-Mills, a cousin of real particle theories, at weak coupling, meaning its particles barely interact. Physicists showed in 1999 that a transition happens there too, but not whether it is abrupt or smooth. In 2003, Ofer Aharony and colleagues showed that the answer comes down to the sign of one number that first appears three loops in, meaning three rounds of ever harder corrections. They did a simpler cousin in 2005. The N=4 version, more than ten times larger, was never finished.
According to the paper, Claude computed that number as b = −1.48 × 10⁻³ λ², where λ measures how strongly the particles interact. Because b is negative, the transition is abrupt, just below the temperature where the gas would otherwise become unstable. Anthropic's reading is that black hole formation is "built into the underlying quantum theory," not only a feature of relativity.
Problem two: how heavy can the lightest black hole be
The second problem is set in a simpler universe with two dimensions of space, which still has black holes. Physicists there use the bootstrap: instead of building the right theory, they rule out theories that contradict themselves. Simeon Hellerman's 2009 modular bootstrap puts a ceiling on the mass of the lightest black hole. Any theory above the ceiling is inconsistent.
Physicists write the ceiling as c/Λ, where c roughly counts the universe's microscopic building blocks. A bigger Λ means a tighter ceiling. Before this work, the paper says, the best value was only known to lie between 8.503 and 12. Computer searches suggested about 9.1 but could not prove it.
The paper gives the exact answer: Λ = 12π² / (π² + 4 arcsinh²1), about 9.1266. It proves two things for large c. A ceiling at that level can be built, and no functional in a broad class can do meaningfully better, so this is the limit of this version of the method. It also shows the method alone can never reach 12, where black holes first appear. Anthropic says Claude spotted a pattern in the numerical results that physicists had missed and brought in a tool from elsewhere in math, a spectral curve. Part of the proof follows the strategy of a sphere-packing proof by an OpenAI model.
How the work was done
Anthropic researcher Hongbin Chen started it in August as a side project. Claude proposed 22 open problems on black holes and the bootstrap, and Chen gave each its own agent. The agents could read each other's progress posts and ran in the background for about six weeks.
For the Hawking-Page result, the paper says almost every number came from two separately written programs. For the bootstrap, Anthropic says an earlier agent failed in 61 parallel attempts, and an agent working on the related sphere-packing problem turned the pattern into a method. Two other agents reached the same answer independently. Anthropic says Claude then wrote the bootstrap paper with minimal guidance, and Hexagon labels both papers, 42 and 131 pages long, as mostly AI-written.
Which Claude model did it
The post names Claude Fable 5.2 once in its body, on the bootstrap result, and in its page title and share text. Its headline and the rest say only Claude. The Fable 5.2 launch post on November 9 had already credited Fable 5.2 with both results. The papers credit an internal Claude research model, and the bootstrap paper adds that the public Claude Opus 5.5 later got the same result on its own. More on the model is in our Fable 5.2 launch article.
Who checked it
So far, the checks are Anthropic's, its models', and unnamed physicists'.
- Hawking-Page. The paper's provenance note says a physicist checked the result against several known special cases. Anthropic says a physicist who had used Claude on the same problem separately reached the same answer by a different method. Different methods, but Claude both times. The paper also reproduces known results, though it says one of those checks "is not a blind test," because a first attempt missed by 10% and led to a fix.
- Bootstrap. Anthropic says the results were verified in Lean, a proof-checking program, in September. The paper does not mention Lean, the post links no Lean files, and Hexagon does not accept Lean code. The paper calls its proof computer-assisted, with some inequalities certified by rigorous interval arithmetic. A physicist checked every main result.
Both papers sit on Hexagon, a repository for math research in public beta, deposited by Anthropic on November 7. Neither is peer reviewed, and no published outside check turned up.
The limits
- Not our universe. Anthropic says neither result describes our universe directly or has practical uses.
- The middle is unproven. The transition is now known to be abrupt at weak and strong coupling. Anthropic calls staying abrupt in between the most natural possibility, but the paper says plainly that a weak-coupling calculation cannot test it. A few terms came from only one program, though the paper says the margin on the sign is far larger than those terms.
- Spinless only. The 9.1266 result applies to the simplest version of the modular bootstrap, which ignores how black holes spin. The paper claims nothing about spin, or about which theories actually exist.
This is a real step for AI in research. The first problem is a long, error-prone calculation people set aside for two decades; the second needed a new idea. But the claims are Anthropic's, checked mostly in-house. The verdict that counts will come from physicists who reproduce the 173 pages.
Frequently asked questions
Which Claude model solved the black hole problems?
Anthropic's November 17 post names Claude Fable 5.2 once in its body, and the November 9 Fable 5.2 launch post credits that model. The papers call it an internal Claude research model, and the bootstrap paper says the public Claude Opus 5.5 later reached the same result.
Has anyone outside Anthropic verified Claude's black hole results?
Not publicly yet. Anthropic says a physicist checked every main result of the bootstrap paper and the bootstrap results were verified in Lean, and a physicist checked the Hawking-Page result against known cases. Neither paper has been peer reviewed.
Do Claude's black hole results apply to our universe?
No. Anthropic says neither result describes our universe directly or has practical applications. Both concern simplified model universes that physicists use to study quantum gravity.
Where can I read the Claude black hole papers?
Both are on the Hexagon preprint server: hexagon 2611.00175 on the Hawking-Page transition and hexagon 2611.00174 on the spinless modular bootstrap, deposited by Anthropic on November 7, 2026.
Sources
- Claude solves two open problems in the physics of black holesAnthropic
- The Hawking-Page transition at weak couplingHexagon (contributed by Anthropic)
- The spinless modular bootstrap at large central chargeHexagon (contributed by Anthropic)
- Introducing Claude Fable 5.2 and Claude Mythos 5.2Anthropic
- The Hagedorn/deconfinement phase transition in weakly coupled large N gauge theoriesarXiv (Aharony, Marsano, Minwalla, Papadodimas, Van Raamsdonk)
- A universal inequality for CFT and quantum gravityarXiv (Hellerman)
- Ten Advances in Mathematics and Theoretical Computer ScienceOpenAI
- About HexagonHexagon Mathematics Foundation




