Claude Code exposes two settings that affect output quality: model selection and effort level. Model selection swaps which set of frozen weights handles a request, determining capability ceiling and per-token cost. Effort level controls how much work Claude does per turn — including files read, tools used, verification steps, and thinking depth — not just thinking time. The guidance is to use default effort levels for most tasks and treat effort as a general preference rather than a per-task dial. When Claude fails, diagnose whether it lacked knowledge (switch to a larger model) or didn't try hard enough (raise effort level). Larger models like Fable excel at complex multi-step tasks and can outperform smaller models on total cost for hard tasks despite higher per-token pricing.
