Anthropic Says Claude Closed a Black Hole Bound Left Between 8.503 and 12
A.I. / news
Anthropic Says Claude Closed a Black Hole Bound Left Between 8.503 and 12
The company's post names one model for the result; the paper's own record names two others, and no outside physicist has reviewed it publicly.

Anthropic said on November 17 that Claude derived an exact bound in the modular bootstrap, a method for ruling out inconsistent theories of 3D quantum gravity, and completed a calculation on the Hawking-Page transition that stalled around 2003. Both results are self-reported by Anthropic. The company says they were formally verified in Lean, a proof checker, and that a physicist checked the main results.
The post, on Anthropic's research page, says an Anthropic researcher named Hongbin Chen asked Claude in August to list open problems. Claude produced 22 candidates. Separate agents then ran in the background for about six weeks.
What the exact bound changes
The second result concerns the mass of the lightest black hole, written M₁, in a 3D gravity theory with central charge c. Hellerman's 2009 proof capped it at c/6. Thomas Hartman, Dalimil Mazáč and Leonardo Rastelli later proved an analytic cap of about c/8.503 in a 2019 paper. Numerical work had pointed to roughly c/9.1.
Anthropic says Claude proved the best this version of the method can do is c/Λ with Λ ≈ 9.1266. The paper's record on Hexagon, which is tagged as contributed by Anthropic, gives the closed form as 12π² divided by the sum of π² and 4 arcsinh²1.
| Result | Year | Ceiling on lightest black hole |
|---|---|---|
| Hellerman proof | 2009 | c/6 |
| Hartman, Mazáč, Rastelli | 2019 | c/8.503 |
| Numerical estimate | not dated | about c/9.1 |
| Claude, per Anthropic | 2026 | c/9.1266 |
| Black hole threshold | n/a | c/12 |
- Hellerman 20096 Λ
- Hartman et al. 20198.5 Λ
- Claude result9.13 Λ
- Black hole threshold12 Λ
Source: Anthropic research post and Hartman, Mazáč, Rastelli arXiv:1905.01319, accessed 2026-11-19
The post says the method alone cannot reach c/12, the black hole threshold. It also says the result does not identify the correct theory of 3D quantum gravity and "may point physicists in the right direction."

Two model names for one result
The post names Claude Fable 5.2 as the model behind the second result. The Hexagon record's provenance statement reads differently. It says an internal Claude research model first obtained the result autonomously, and that the publicly available Claude Opus 5.5 later reached the same answer independently.
The two accounts may describe different runs. Neither explains which model produced the proof that was verified. Anthropic did not say in the post, and The Terminal has not seen a statement that reconciles them.
The dates also differ. Hexagon lists version 1 of the paper as deposited on November 7, ten days before the post.
What has not been checked
The physicist who checked the main results is not named in the post. The record tags the text as "Primarily AI-generated text" and marks "Human understanding: some parts." It lists no individual authors.
The first result is also partial. Claude found the Hawking-Page transition is abrupt at weak coupling, as Hawking and Page found in 1983 using general relativity. The post calls it the most natural possibility that the transition stays abrupt at every coupling in between, and says that is not proven.
The post describes no review by the wider physics community. The post also says an earlier Claude attempt at the bootstrap problem failed in 61 parallel tries.
The record on Hexagon is public under a CC BY 4.0 licence, which means other physicists can check the functional against Λ ≈ 9.1266. Anthropic's earlier work on Fable 5.2 and protein sequences came with a benchmark gap rather than a proof, and the company's IPO, priced at $83 a share, puts every claim it publishes under closer reading. Independent replies on the paper's record would be the next signal.
Sources
More in A.I.
- 01Tencent Hy4-flash Card Says Weights Are Out; Launch Coverage Says PlannedThe Hugging Face page lists 220B parameters and a custom licence, while launch coverage gives 224B and a future release, at $0.070 per million input tokens.
- 02Four Students Sue Omnilert and Baltimore County Over the Chip-Bag Gun Alert at Kenwood HighThe Nov. 17 suit targets a $2.6 million school contract, and the district's own timeline shows staff cancelled the alert 1 to 2 minutes after it fired.
- 03Google Ranks Gemini 4 Neon Third on Artificial Analysis; the Site's Own Page Says SixthGoogle's Nov. 18 launch post sells Neon as fast, but Artificial Analysis clocks 21.5 minutes per task against a 12.1-minute median.
- 04Gemini 4 Carbon Takes 36 Minutes a Task; Kernel Counts 7x the Steps of AstraArtificial Analysis puts Carbon at 36.0 minutes per Intelligence Index task against a 12.4-minute median, and a Kernel study traces most of the gap to extra steps.