Claude black hole physics: a woman in a mustard sweater looks up at a wall of dark cabinets, two of them open and glowing orange and violet (AI image)

Illustration

Developments3 min read

Anthropic says Claude Fable 5.2 solved two open black-hole problems

Anthropic said in a post on 17 November that Claude Fable 5.2 solved two open problems in theoretical physics about black holes. Several Claude agents ran in the background for about six weeks and solved two of 22 candidates, the company said; the preprint for the first problem does not mention peer review.

Key facts

  1. Anthropic published the research post on 17 November 2026, saying Claude Fable 5.2 solved two open problems in black-hole physics.
  2. Agents solved two of 22 candidates Claude proposed in about six weeks; one result was formally verified in Lean in September.
  3. The Hawking-Page preprint is licensed CC BY 4.0 on Hexagon, tagged as primarily AI-generated text, and does not mention peer review.

Manfred „Luigi“ LugmayrBy Manfred „Luigi“ Lugmayr

Researched and written with AI, to Luigi's brief, positions and editorial rules. Automatically checked against the sources. How I work →

Anthropic researcher Hongbin Chen began testing Claude on such questions as a side project in August, according to the post from the company's Science team. He asked Claude to list open problems in theoretical physics around black holes, and Claude returned 22 candidates a few hours later. Chen set up a separate agent for each problem; the agents could read one another's progress posts and ran in the background for about six weeks with no special set-up. Two problems were solved, or just over 9% of the candidates (our calculation).

The Hawking-Page transition at weak coupling

The first problem concerns the abrupt switch from a hot gas of particles to a black hole in a holographic model universe, described by Hawking and Page in 1983. Whether that transition stays abrupt at weak coupling hinges on a quantity Aharony and colleagues identified in 2003, according to Anthropic. The calculation involves many thousands of diagrams and was never finished for the theory N=4 super Yang-Mills. Claude ran a small group of agents, all copies of itself, and had almost every step done twice by independent agents. According to Anthropic, the transition is abrupt at both weak and strong coupling. That it stays abrupt in between is the most natural possibility, but it has not been shown.

The accompanying preprint, “The Hawking-Page transition at weak coupling”, on the Hexagon platform is dated 7 November and licensed CC BY 4.0. It lists no authors but names Anthropic as the contributor, and it is tagged “Primarily AI-generated text”. A physicist checked the result against known special cases, the page says; it does not mention peer review.

A ceiling for the lightest black hole

The second problem concerns the modular bootstrap in three-dimensional gravity, a method that rules out inconsistent theories instead of solving one directly. Hellerman showed in 2009 that the lightest black hole weighs at most c/6 up to small corrections, where c roughly counts the microscopic building blocks. Hartman, Mazáč and Rastelli proved c/8.5 in 2019, while numerical work pointed to about c/9.1. Claude used a spectral curve to prove that the ceiling is c/Λ, with Λ about 9.1266, and that this ceiling cannot be lowered with the simplest version of the method.

Anthropic describes the route this way: one agent noticed a pattern in numerical results while surveying the literature, and another, working on sphere packing, turned it into a method. An earlier Claude agent had failed on the problem in 61 parallel attempts; two further agents independently reached the same answer. One half of the proof follows the strategy of a sphere-packing proof that an OpenAI model recently produced, Anthropic says. The result was formally verified in Lean, a proof checker, in September, and a physicist checked every main result.

Anthropic itself points to limits: neither result describes our universe directly or has practical applications. The bootstrap version ignores black-hole spin and cannot reach c/12, where black holes first appear, on its own. Quantum gravity is not settled.

OpenAI recently reported that agents finish 46% of its multi-hour research tasks and that its models helped with AMD's GPU design; Google Labs opened the waitlist for Stitch Loop, an agent that works around the clock.

Analysis

For companies the set-up matters more than the physics. Anthropic did not use a chat window; it handed 22 bounded tasks to separate agents whose output can be checked: by independent duplicate calculation, by a proof checker and by a human expert. The same logic suits a first agent project inside a company: one clearly defined job that ends in a result someone can judge. In marketing, that could be a reactivation campaign the agent prepares as a package for review instead of launching it.

The hit rate tempers the picture. Two of 22 candidates is just over 9%, and the 61 failed parallel attempts by an earlier agent show how much waste such a run produces. Anthropic does not give compute or cost. Teams that let agents run for weeks should decide in advance who reviews the output and log the share of usable results per task. A starting point is the guide AI agents in marketing.

  • Share

Got AI news? Submit news →

FAQ

What did Claude solve, according to Anthropic?

Two open black-hole problems: whether the Hawking-Page transition stays abrupt at weak coupling, and an exact ceiling for the lightest black hole in 3D gravity (c/Λ, Λ about 9.1266).

Are the results peer reviewed?

The Hexagon page for the Hawking-Page preprint does not mention peer review. Anthropic says a physicist checked the main results, and one result was formally verified in Lean.

Source

Anthropic: Claude Fable 5.2 solves two open problems in the physics of black holes (17.11.2026)

More sources

  • Hexagon – Preprint-Seite: Anthropic als Einreicher, CC BY 4.0, „Primarily AI-generated text“, kein Gutachterverfahren erwähnt.
  • Hexagon – Zweites Preprint, im Anthropic-Beitrag verlinkt; Angaben dazu stammen aus dem Anthropic-Beitrag.
  • arXiv – Hintergrund: bewiesene Grenze c/8,5.

Black Friday 2026

Black Friday deals on AI tools, straight to your inbox.

I collect the AI tool deals worth having and send them in the AI Briefing, along with the AI news that matters, every Wednesday and Friday. Five minutes to read.

  • Black Friday deals on AI tools, collected for you
  • The key AI news, briefly explained
  • Wednesday and Friday, unsubscribe in one click

In the AI Briefing right now

Image for the story: Superlist Black Friday: Super plan costs $99.50 in year oneDealSuperlist Black Friday: Super plan costs $99.50 in year one
Image for the story: Anthropic says Claude Fable 5.2 solved two open black-hole problemsNewsAnthropic says Claude Fable 5.2 solved two open black-hole problems
Image for the story: Google tests Chrome's AI agent against attacks in a replica of the webNewsGoogle tests Chrome's AI agent against attacks in a replica of the web
Blonde woman with glasses smiling as she scrolls on her phone in an office

Manfred “Luigi” LugmayrFrom Manfred “Luigi” Lugmayr: 30 years in marketing, seven of them as CMO at burgerme.