Anthropic Says Its New AI Finishes Math Too Long for People to Do
Its examples include black-hole calculations, gene-therapy shells, and a computer-checkable version of Fermat's famous proof.

The AI Daily Brief: Artificial Intelligence News and Analysis · Fable 5.2 Is the Best AI Ever. So Why Am I Not Switching Yet?Anthropic says its new AI system, Fable 5.2, can finish long, exact calculations that are too lengthy for people to complete by hand or with standard software. These are the company's own claims about its tests, not independently checked results.
As one example, the company says the system solved two open, calculation-heavy problems about black holes in theoretical physics. It says the work meant grinding through derivations too laborious for human scientists using usual tools.
In its own tests, the company also says the system accurately predicted the shape of 28 of 34 protein shells used to deliver gene therapy, called capsids. Anthropic says that kind of prediction could help biologists design new delivery shells faster and more cheaply.
For math, the company rewrote Fermat's famous theorem, proved in the 1990s, into a version written in strict logic that a computer can check step by step. That job took 30 hours using 1,000 copies of the AI working on parts of the task at once, while Opus 5.5, an earlier Anthropic model, reached 26% in 18 hours and then stalled.
Questions people ask
If mathematicians already proved Fermat's Last Theorem, what does a checkable version add?
A formal version is rewritten in strict logic that checking software can verify line by line. For Fermat's theorem, already proved in the 1990s, that process can expose hidden gaps and lets computers confirm every step.
Why do gene-therapy researchers care about predicting capsid shapes?
Gene therapy often uses modified viruses to carry replacement genes, and the capsid is the outer shell that holds that cargo. Its shape affects where the therapy goes in the body and whether the immune system attacks it, so faster shape prediction can speed up design.
How can grinding through calculations solve an open physics problem?
A theory can stall when scientists cannot finish the equations needed to test it against reality. Anthropic describes its black-hole results as that kind of laborious calculation, where the AI worked through very long derivations.
