Anthropic unveils Claude Fable, a narrative-native frontier model

Anthropic today released Claude Fable, a new frontier model trained explicitly for long-form narrative reasoning — the first from a major lab to treat storytelling as a first-class evaluation axis rather than a decorative side-effect of chat.

In internal benchmarks the model holds a coherent 250,000-token plot with named-entity continuity above 98%, and — more surprisingly — outperforms Opus 4.8 on multi-step tool use when the tools are described as characters in a fable rather than as JSON schemas.

“The interesting result is not that Fable writes better fiction. It is that narrative structure appears to be a compression format the model already reasons in.”

Pricing is aggressive: $2.40 per million input tokens, $12 per million output — roughly a third of Opus 4.8 — with a 1M-token context window generally available from day one. Anthropic is positioning Fable at agent builders, screenwriters and, pointedly, the games industry.

The release lands the same week OpenAI's S-1 pointed to 'narrative agents' as a growth vector. Whether Fable is a genuine capability step or a very expensive marketing exercise will be settled by the third-party evals over the next fortnight.