Claude Code optimization: choosing models and effort levels

Anthropic engineer Lidia Halley released a guide on balancing costs between Fable, Opus, and Sonnet. The effort setting is passed as part of the prompt to control planning depth. When debugging at high effort, the model can rewrite its plan on the fly if an initial check yields immediate results.

Halley suggests a simple heuristic: if the model fails, determine if it lacks knowledge or effort. Increase effort for lack of diligence, but switch to a larger model for lack of knowledge. For simple tasks, a smaller model at high effort is often more efficient than a larger one.

Internal tests show that Opus 4.8 at a standard effort level outperforms Opus 4.7 while consuming the same number of tokens.