Anthropic Releases Claude Fable 5.2: 1,000 Agents Formalized Fermat's Theorem in 30 Hours

Anthropic has released two models at once — Claude Fable 5.2 and Mythos 5.2. Under the hood they share a single base model, and the difference lies in the safeguards. The loudest demonstration was this: a thousand agents formalized the proof of Fermat's Last Theorem in Lean in 30 hours. It sounds like a pop-science stunt, but behind it stands a fairly down-to-earth engineering story.

What Was Shown in Practice

Lean is a language in which a compiler, not a reviewer, checks the proof. Any gap in reasoning simply won't compile. That makes formalization enormous, tedious and extremely strict work, similar to refactoring a giant codebase with a rigid type system. The previous research version of Claude needed 11 days for it, while a thousand agents of the new model needed 30 hours. For a developer, the theorem itself isn't what matters, but the pattern: a big task is cut into pieces, hundreds of agents take the pieces in parallel, and an automatic verifier weeds out the defects.

Anthropic's other claims follow the same vein. Fable 5.2 solved two open problems related to black hole physics. The model designed three working devices — with printed circuit boards, enclosures and software. And more than 90% of the changes that land in Anthropic's own core code are now written by Claude with human involvement. It's a strong claim, but it is the company's data about itself, with no independent verification yet.

Price, Speed and Two Versions

The new generation is up to 30% faster than the previous one. Pricing: $8 per million input tokens and $40 per million output tokens. There is a Fast Mode — up to 2.5 times faster, but at $16 and $80 respectively. The arithmetic is simple: the speedup costs exactly twice as much, so the mode makes sense where latency matters — interactive assistants, live IDE agents — rather than in overnight batch jobs.

The versions differ in access. Fable 5.2 is open to a broad range of users and is already available in Claude, via the API, and on AWS, Google Cloud and Microsoft Azure. Mythos 5.2 is for vetted organizations in cybersecurity and biomedical research. According to Anthropic's internal test, it finds 10.2% more vulnerabilities than Opus 5.5. The test is internal, so the figure should be treated as a claim, not an independent benchmark.

What Developers and Product Managers Should Do

First, recalculate the economics of your agent pipelines. Output tokens cost five times more than input tokens, so agents that "think" and write a lot will become the main line of spending. Set limits on output length and cache everything you can.

Second, borrow the idea from Lean: give your agents a strict automatic verifier. In code, that means tests, types and linters; in data, schemas and validators. The stricter the check, the safer it is to hand agents parallel work.

Third, don't take the numbers at face value. "Up to 30% faster," the 10.2% gain, 90% of code — these are all the vendor's words. Run Fable 5.2 on your own tasks and compare the cost of a solved task, not the cost of a token.

Finally, if you work in security or biomedicine, find out how organization vetting for Mythos access works. Everyone else will be fine with Fable: the base model is shared, and only the restrictions differ.