Claude Fable 5.2 and Mythos 5.2 explained: one model, two sets of safeguards, and a lower price

They are Anthropic's largest models, meant for long, hard technical work: big software migrations, research, and tasks that run for hours. Anthropic's announcement says it plainly: "Claude Fable and Claude Mythos are the same underlying model."

So why two names? The difference is the safeguards. Fable 5.2 runs with safety classifiers that block high-risk cybersecurity work, such as penetration testing and exploit generation, and most research biology tasks. Because of those blocks, Anthropic can offer Fable to all users. Mythos 5.2 is the same model without those blocks, so it is only given to vetted organisations doing defensive security or biomedical research.

When Fable blocks a question, you are not left with nothing. The announcement says a blocked cybersecurity query gets an answer from Claude Opus 4.8 instead, and a blocked biology query gets one from Claude Opus 5.

Less often than you might think, and Anthropic says so itself. Its advice is to start most coding and knowledge tasks on Claude Opus 5.5, which it calls its "strongest daily driver", and only move up to Fable or Mythos when the work is large and open-ended. For simpler work, you move down to Claude Sonnet 5.5 or Claude Haiku 5.5.

Anthropic also describes two ways to use Fable without sending it every request. As an advisor, a smaller model does the work and asks Fable for help only on the hard parts. As an orchestrator, Fable plans the work and hands the simpler pieces to smaller models. Both patterns keep most tokens on cheaper models, which matters when output costs $40 per million tokens.

This is the same thinking you use for any expensive resource in a system: route by difficulty, and pay for the strongest option only where it changes the result. Test it on your own tasks before you switch, because a benchmark average does not tell you how a model does on your code.

Read next: Model routing and cascades: sending each request to the right model

The interesting design choice here is not the benchmark scores. It is releasing one model twice and letting the safeguards, not the weights, decide who gets which capability. A classifier sits in front of the model, and a blocked request falls back to an older model rather than failing. That is a routing and fallback design, and it is worth understanding for any product that puts a model in front of users.

For most teams, the practical question is cost. At $40 per million output tokens, Fable 5.2 is a model you call for the hardest step of a task, not for every step. Measure where a stronger model actually changes your results, route only those requests to it, and keep everything else on a cheaper model.

Read next: The AI Engineering course: models, routing, evals and cost, from the ground up