Anthropic introduced Claude Fable 5.2 and Claude Mythos 5.2, reporting improvements in coding, research and reasoning. In an experiment described by the company, about 1,000 Fable agents completed a Lean formalization of the proof of Fermat’s Last Theorem in 30 hours with minimal human intervention.
What the experiment achieved
Formalization is not a first-ever solution of the theorem. It translates a proof into a language in which a computer can check its logical steps. Anthropic says an earlier research version of Claude required 11 days, a purpose-built agent harness and guidance from a researcher. The new experiment supplied short instructions and intermediate theorem statements for measuring progress, while the agents built their own coordination mechanisms.
The comparison involved a single run per model. That does not establish a repeatable success rate, typical cost or performance on unrelated tasks.
How Fable and Mythos differ
Both versions use the same underlying model. Fable has stronger safeguards against misuse in cybersecurity and biology. Mythos access is intended for vetted organizations through Anthropic’s cyber and life sciences verification programs. Broader Fable availability does not mean every activity is permitted or that usage is free.
Fable 5.2 is available through Claude Platform and platforms including AWS, Google Cloud and Microsoft Azure. Standard API pricing is $8 per million input tokens and $40 per million output tokens. Fast mode costs $16/$80 and offers up to 2.5 times faster token generation; this is not a guarantee that every task finishes proportionally faster.
The practical lesson
The interesting development is the combination of parallel execution and formal checking, not simply the number of agents. For businesses and development teams in Israel, the useful question is which outputs can be checked through software tests, rules or formal verification. Lean is designed for mathematical proofs, not as a universal certificate of AI reliability.
Anthropic itself recommends starting most everyday coding and knowledge tasks with Opus 5.5, and selecting Fable according to task complexity. These results are attributed to the company, not independent EzraTech testing.
Source: https://www.anthropic.com/claude-fable-and-mythos-5-2

