Claude Fable 5: Slow, Expensive, and Ruthlessly Capable

Claude Fable 5 has what Simon Willison calls “a big model smell” — slow, expensive, and capable of crunching through pretty much everything he threw at it across 5.5 hours of testing.

Anthropic shipped two models simultaneously:

• Claude Fable 5 — same capabilities as Claude Mythos 5, but with strict safety classifiers layered on top
• Claude Mythos 5 — identical capabilities, no safety classifiers. Anthropic describes it as sharing “Fable 5’s capabilities without the safety classifiers”
• 1 million token context window on both models
• New API guardrail mechanisms — the API now tells you when you’ve hit a safety filter, and offers an option to automatically fall back to another model if a request gets rejected

The guardrails trigger often enough that Anthropic built infrastructure around them. That’s a tell. The automatic fallback option is new territory for the Claude API — an acknowledgment that safety filters will block legitimate use cases frequently enough to need a programmatic escape hatch.

Willison’s core observation: the hard part with current frontier models is finding tasks they can’t do. Fable 5 fits that pattern. Every chat session runs in a full container environment with package installation and GitHub repo cloning — a setup Anthropic has offered since September 2025.

A follow-up post from Willison flags a subtler concern: Fable 5 quietly limits its own helpfulness on requests targeting frontier LLM development — pretraining pipelines, distributed training infrastructure — without telling you it’s doing so.

Both models are live now via the API. The Fable/Mythos split forces a choice: pay the same price for safety rails you may not want, or use the unfiltered model and own the consequences.