Anthropic Unveils Claude Fable 5.2: Fermat's Theorem in Lean in 30 Hours Instead of 11 Days

When a lab says "AI now does science," the first question is simple: what exactly did it do, and who checked it? Anthropic has released Claude Fable 5.2 and Mythos 5.2 and brought four claims at once. They carry very different weight, and lumping them into a single "breakthrough" headline is a mistake.

What Anthropic Actually Claimed

According to the company, Fable 5.2 formalized a proof of Fermat's Last Theorem in Lean in 30 hours. The previous version needed 11 days for the same task. In a second experiment, Fable coordinated more than a thousand AI agents, and over eight days they wrote a prototype operating system with its own kernel, drivers, and graphical interface.

The other two claims concern science. Anthropic reports solving two open problems in theoretical physics related to black holes and quantum gravity. The Mythos model, it says, proposed a new method for modeling viral protein shells.

Where the Real Signal Is

The most valuable story for practitioners is the Lean one. Lean is a system in which a proof is checked not by a reviewer but by a compiler. Either the chain of reasoning holds together, or the file doesn't build. Formalizing a known proof means translating human text, with phrases like "it is obvious that," into a strict machine-readable form. It is enormous routine labor, and cutting it from 11 days to 30 hours is a nearly ninefold speedup.

Something else matters too: this kind of result can be verified without taking the lab's word for it. If Anthropic publishes the source files, anyone can run them through Lean. It is the same logic as tests and a compiler in ordinary development: the stricter the automatic checker, the safer it is to let an agent loose on long-running work.

The story of the thousand agents is interesting not because of the OS itself but because the model acts as a conductor. The question is no longer how well it writes a function, but whether it holds a plan, hands out tasks, and assembles the result over a week of work. The post says nothing about cost, the number of errors, or how far the "prototype" is from a working system. Without those figures, the scale cannot be judged.

What Remains in Question

The author of the original post himself notes important caveats. The AI did not prove Fermat's theorem anew: the proof has long been known, and it only translated it into formal form. This is an engineering achievement, not a mathematical discovery.

The physics calls for even more caution. The results are so far presented as preprints, meaning they have been published rather than peer-reviewed. Most of the experiments were run by Anthropic itself. The material contains no independent reproductions, no comparisons with existing methods, and no metric figures for the viral protein shells. For now this is a claim, not an established fact.

What to Do With This

Developers. Take the pattern, not the brand. Where an agent's output is checked by a machine (Lean, tests, types, simulation), long autonomous tasks are already realistic. Where verification boils down to "looks about right," it is too early to delegate that much.

Product managers. Look at task duration, not five-minute benchmarks. The drop from 11 days to 30 hours suggests that models are improving quickly over long horizons. Factor that into planning, but test it on your own cases.

Founders. Don't build a roadmap on a press release. Wait for primary data, independent reproductions, and access terms: the post names neither pricing nor API availability. For now you can calmly note that agentic systems have learned to work for days rather than minutes. But claims about "solved physics problems" are best set aside until someone other than the authors checks them.