Blog · Zakhar Garan
Fable, Opus, Sonnet or Haiku? How not to overpay for AI without losing quality
Anthropic has released Claude Fable 5.2, so which model should you pick for which task?
- Opus 5.5 is recommended as the default for complex code, at medium effort. If the result is weak, raise the effort level first and only then switch models: it is cheaper, and the cache is preserved.
- Fable 5.2 is for large, fuzzy tasks where a mistake is expensive, such as rewriting the architecture of a whole product. Its tokens cost twice as much as Opus, but it plans on its own and tests as it goes.
- Sonnet suits well-defined tasks where speed and budget matter most, such as a bug that is easy to reproduce. It costs half as much as Opus.
- Haiku is for large volumes of routine work: sorting customer requests or writing short summaries.
You can combine models. For example, Fable can lead the migration of a whole project to a new framework version and hand off updates of individual packages to Sonnet subagents.
Are you already combining different models in your work?