Claude finished a two-decade-old calculation. Both preprints are still v1.
Anthropic says Claude Fable 5.2 solved two open problems in black hole physics. One closes a 2003 calculation, the other is harder to price from outside.

Anthropic published a research post on November 19 claiming that Claude, which the post calls Fable 5.2, solved two open problems in the physics of black holes. The preprints are real and public: both load, both are credited to Anthropic, both went up on a repository called Hexagon on November 7 as version 1 under a CC BY license. Twelve days separate the papers from the announcement. Nothing about the claim is fabricated, and nothing here has been refereed.
The problem that arrived with a target already painted
The first result has an advantage the second does not: somebody else set the target, in 2003, before the model existed.
In 1983, Stephen Hawking and Don Page showed that a model universe full of hot particles turns into a black hole through an abrupt phase transition, much as water boils. Their calculation needs general relativity, which holds only for a large universe. Shrink the model toward the scale where spacetime stops having a definite shape and it stops working. In 1999, Bo Sundborg and then a group including Ofer Aharony showed a transition still exists on the quantum side, where the math does work. Nobody could say whether it was abrupt or smooth.
Then in 2003, Aharony, Marsano, Minwalla, Papadodimas and Van Raamsdonk worked out what would settle it: the sign of one coefficient, the quartic term in the effective action of a certain loop, which first appears at three loops. Computing it meant thousands of diagrams plus a set of infinities that must be tamed first. The group did a smaller version in 2005 and never finished the larger one.
That is the specification Claude was handed, which is why this claim is easy to price. The first preprint reports b = −1.48×10⁻³ λ² in terms of the 't Hooft coupling. Negative means first order, an abrupt jump. It also places the answer slightly below the Hagedorn temperature and reproduces the two-loop value known independently from integrability. Each of those is a place where a wrong sign or dropped factor would surface as a contradiction with published work.
| Hawking-Page at weak coupling | Spinless modular bootstrap | |
|---|---|---|
| Question | Does the transition stay abrupt? | How light must the lightest black hole be? |
| Prior state | 2003 setup, one number uncomputed | Proven 8.503, numerics near 9.1 |
| Claimed result | First order, b = −1.48×10⁻³ λ² | Λ ≈ 9.1266, exact form in the paper |
The second claim is a different kind of object
The second preprint concerns 3D quantum gravity and the modular bootstrap, which does not build a correct theory. It rules theories out by showing they contradict themselves, and in doing so sets a ceiling on how heavy the lightest black hole can be, written c/Λ. A larger Λ means a tighter ceiling.
Before this paper the slope Λ was only known to lie between 8.503 and 12, the upper end being the BTZ black hole threshold, where the bound would stop constraining anything. Numerics had been pointing near 9.1 without proving it. Claude's paper closes the gap: Λ is exactly 12π²/(π² + 4 arcsinh²1), about 9.1266.
The number landed essentially where the numerics already pointed. The contribution is not a surprising value but a proof, plus a closed form nobody had written down, obtained by importing a spectral curve from elsewhere in mathematical physics. The post describes Claude as spotting a pattern physicists had missed. The narrower gloss: the pattern was visible in the numbers; the proof took an idea from outside the subfield.
The optimality claim needs reading too. The post says the ceiling "is optimal: it's the lowest this version of the modular bootstrap can ever go." The abstract is more careful. No linear functional, in a class containing all finite combinations of derivatives, beats c/Λ by more than a correction of order c-one-third times a logarithm. That is a statement about a specific method, not about quantum gravity. The paper also conjectures the coefficient of that correction, supported by numerics up to c = 2048. A conjecture inside a proof paper is normal; it should not be reported as proved.
How the work was actually run
The provenance deserves more attention than it usually gets. An Anthropic researcher, Hongbin Chen, asked Claude as a side project for a list of interesting open problems in theoretical physics. It returned 22 candidates hours later. Separate agents were set up per problem, able to read each other's progress posts, and ran in the background with no special setup for about six weeks.
Some of that arrangement produced genuine checks. Three separate agents independently reached the same bootstrap answer; an earlier agent had failed at that problem across 61 parallel attempts. On the phase transition, a physicist working independently with Claude reached the same conclusion by a different route, and a colleague noticed the agent had got there first. Anthropic says the results were formally verified in Lean, a computerized proof checker, in September, and that a physicist checked every main result.
The repository adds a disclosure the post does not mention. Above the abstract of both papers, Hexagon prints a provenance label reading "Primarily AI-generated text" and, for human understanding, "some parts." That is more candid than most vendor announcements, and it is the most useful fact for anyone deciding how much weight to put on either paper.
Where the checking bites, and where it needs a specialist
The phase transition result is checkable in principle by anyone willing to redo a three-loop computation a research group abandoned. The target was set externally in 2003, the answer is a sign, and the paper reports that almost every number in it was obtained in two independent ways. If the sign is wrong, the paper is wrong, and that is decidable.
The bootstrap result resists that treatment. Its novelty is a technique import, so the work depends on whether the tool was applied correctly and whether the class of functionals considered is the right one. Checking that from outside would take a specialist in conformal field theory and modular invariance working through 85 references, and even then would likely mean rederiving the spectral curve rather than verifying a number. Lean helps only partly: it confirms the logic once the statement has been formalized, and the formalization is itself an act of translation nobody outside Anthropic has yet audited.
The first result will either survive contact with the literature or quietly fail, and the failure mode is legible. The second could be correct, could carry an over-scoped optimality claim, and most readers could not tell the difference.
What the post says about its own limits
Anthropic is unusually direct about the boundaries. "Neither result describes our own universe directly, and neither has practical applications," the post states, and later: "Neither result settles the question of quantum gravity, or even comes close to doing so, but each pushes physics forward." It describes the first as a known calculation too error-prone to finish by hand and the second as requiring a new idea, and notes that this version of the bootstrap cannot reach c/12 on its own.
Those sentences should travel with the headline. These are contributions to a theoretical program, not a breakthrough in understanding gravity.
The pattern worth watching
The quantum circuit paper Anthropic published in October had a checker: a formula that scored each candidate design, so any third party could audit the arithmetic. This is the same lab making a much larger claim with a much weaker instrument. One of the black hole results has a target defined by strangers and an answer that reduces to a sign. The other needs a specialist to price.
Independent refereeing will do more for both than any commentary. Until then, the fair statement is that Anthropic's model produced two preprints of real apparent substance: one closes a calculation the field left open in 2003, the other proves a ceiling the numerics had already guessed at, and the problem set was chosen by the same organization that benefits from the problems being solved. On Hacker News the post drew 550 points and 41 comments, wide sharing with comparatively little argument underneath.


