Anthropic Says Fable 5.2 Closed Two Open Physics Problems — Here Is the Evidence

Anthropic announced on 9 November 2026 that Claude Fable 5.2 closed two open problems in theoretical physics, both in holographic duality — the correspondence that links gravitational universes to quantum theories without gravity in them. The claim is unusually specific, which makes it unusually checkable. Here is what each problem was, what the answer is, and exactly how far the verification goes.

Problem one: do black holes still form abruptly in a very small universe?

In 1983, physicists showed that as one of these gravitational universes heats up, a black hole appears abruptly rather than gradually — a sharp transition, closer to water reaching its boiling point than to a kettle warming. The argument worked only for universes large enough for general relativity to apply cleanly.

Whether the same sharp transition happens in a very small universe was an open question. A method for testing it was laid out in 2003 and then left unfinished, because carrying it through was too laborious to be worth anyone’s time.

Anthropic says Fable 5.2 completed that calculation and found the transition is abrupt in the small case too. The result turned on the sign of a single coefficient that first appears at three loops — which is the kind of detail that makes the claim credible rather than vague. It is a specific quantity with a specific sign, and anyone with the expertise can go and check it.

Problem two: the exact limit of the spinless modular bootstrap

The modular bootstrap is a technique physicists have used for more than fifteen years to rule out candidate theories of quantum gravity in simplified universes. It works by establishing a bound: anything beyond the bound cannot be a consistent theory.

The trouble was that nobody knew where the bound actually sat. It was known only to lie somewhere between 8.503 and 12. Anthropic says Fable 5.2 found the exact value: approximately 9.1266, written up in a 131-page paper.

How much of this has been checked, and by whom

This is the part most coverage skipped, and it matters more than either result.

Read the third row carefully. In both cases the independent confirmation came from another Claude model. That is a real and useful check — two different methods agreeing is meaningful, and a model reproducing a result it was not shown is not nothing. But it is not independent verification in the sense physics normally means, because a shared training lineage can produce a shared mistake. The human checking is real too, and the physicists who did it have not been named.

Neither paper has been through peer review. Both were deposited by Anthropic on a preprint server, which is the ordinary first step for theoretical physics and carries no vetting by itself.

Why this is a different kind of claim from the usual one

“AI solves hard problem” is usually unfalsifiable — a benchmark score, a demo, a paraphrase. This is not that, for three reasons:

• The problems are named and were open before. The 2003 method existed and nobody finished it. The bootstrap bound was a known gap with known endpoints.
• The answers are numbers. Roughly 9.1266, and the sign of a specific three-loop coefficient. Either they are right or they are not, and the field can tell which.
• The working is published. A 131-page paper with every main result checked by a human is a thing other physicists can attack. Anthropic has made this falsifiable on purpose.

That is the meaningful shift. Not that a model produced an answer, but that it produced one specific enough to be wrong in public.

What would settle it

• Peer review on both preprints. Until then these are claims with working shown, which is a respectable status and not the same as established results.
• A named physicist outside Anthropic confirming the bootstrap value. The number is checkable by anyone in the field with the right tools, so this should arrive relatively quickly if the result holds.
• Replication with a non-Claude system, or by hand. This is the gap in the current verification chain and the easiest one to close.
• The three-loop coefficient confirmed independently. A sign is the simplest possible thing to check and the simplest to get wrong.

None of that diminishes the result. It describes where it sits: two concrete, published, checkable claims from a company that has an obvious interest in them being true, released the same week as the model that produced them.

The rest of what Fable 5.2 was reported to have done

The physics was one item on a longer list from the same announcement. Also claimed: a team of roughly 1,000 Fable 5.2 agents formalising Fermat’s Last Theorem in 30 hours; an operating system that boots on a laptop, built by over 1,000 subagents across eight days; and three electronic devices designed with Diode — boards, enclosures and software — all working without electrical modification. The full launch breakdown, with pricing and every benchmark, is here, and we compared Fable 5.2 against Opus 5.5, GPT-6 Astra and Gemini 4.

Frequently asked questions