Anthropic’s Claude Fable 5.2 finishes a 23-year-old black hole calculation

An AI model just finished homework that physicists started in 2003.

Anthropic’s Claude Fable 5.2, formally announced on November 9, 2026, completed a black hole calculation that had gone unresolved for 23 years. The result: in very small universes, black holes don’t ease into existence. They appear suddenly.

What the model actually worked out

The breakthrough came during prerelease testing, before the model was publicly available. Fable 5.2 completed the calculation that began in 2003 and showed that black hole formation in tiny universes happens through a first-order phase transition.

If that phrase sounds intimidating, picture a pot of water on the stove. Nothing visible happens for a while, then the water abruptly starts to boil. A first-order phase transition is that kind of sharp, all-at-once change, and the research describes this black hole behavior as akin to boiling.

The physics framework behind the result is holographic duality. Under that approach, the dynamics of gravity in one description line up with a quantum theory that contains no gravity at all.

Fable 5.2 didn’t stop at one problem. It also tackled a second open question in black hole physics, establishing the precise limit of existing techniques used to rule out various quantum gravity candidates in simplified universes.

The research describes the model as having addressed both problems autonomously. The findings are laid out in two preprints released after the announcement.

Preprints, for the uninitiated, are papers shared publicly before formal peer review. So the scientific community now gets its turn to check the work, which is how physics is supposed to operate whether the author is a postdoc or a language model.

The models, the price tag, and the gatekeeping

Fable 5.2 arrives alongside a sibling, Claude Mythos 5.2. Anthropic developed both with a focus on technical work: coding, reasoning, and scientific problem-solving.

Pricing is set at $8 per million input tokens and $40 per million output tokens. Tokens are the chunks of text a model reads and writes, so input is what you feed it and output is what it hands back. The models are available through paid plans.

Mythos 5.2 comes with a velvet rope. Access is restricted to verified organizations, particularly those working in cybersecurity and biomedical research.

Why a physics result matters for an AI launch

A completed calculation on a problem that sat open since 2003 is a different kind of evidence. It is a concrete research output, written up in preprints, that other physicists can scrutinize line by line.

Anthropic surfaced the black hole result from prerelease testing and paired it with the launch announcement.

What this means

Two preprints now sit in public view, and their value depends on whether the physics community confirms the results.

The second result may prove more practically useful than the headline-grabbing first one. Mapping the exact limits of techniques used to eliminate quantum gravity candidates helps physicists decide where to stop pushing old methods and where new ones are needed.

At $8 per million input tokens and $40 per million output tokens, output costs five times as much as input. Workloads that generate long reasoning chains, like multi-step derivations, will feel that difference most.

By gating Mythos 5.2 to verified organizations in sensitive sectors, Anthropic is testing a tiered model of distribution: broadly available tools on one side, restricted capabilities on the other.

The thing to watch next is the reception of those two preprints. A 23-year-old problem closed by software is a striking claim. Its lasting significance will be decided by the physicists who now get to check the math.