Anthropic Breaks Down Claude Model Selection and Effort Levels in Claude Code
Key Takeaways
- ▸Model selection sets capability range; effort level controls how much work Claude invests per request (file reads, verification, task depth)
- ▸Use smaller models for routine tasks, larger models for complex/ambiguous tasks—start with default effort and tune by work type, not per-task
- ▸If Claude fails despite having context, upgrade the model; if it fails by skipping files or bailing on refactors, increase effort level
Summary
Anthropic has published guidance explaining how developers should choose between different Claude models and effort levels when using Claude Code. The company clarifies that model selection determines the model's overall capability and knowledge base, while effort level—distinct from simple 'thinking time'—controls how comprehensively Claude approaches a request, including the number of files it reads, how much it verifies, and how far it pushes through multi-step tasks before checking back with the user.
The guidance recommends matching task complexity to model capability: use smaller models like Claude Sonnet for routine tasks and larger models like Claude Fable for complex or ambiguous work. For effort level, Anthropic suggests tuning it as a general preference based on work type rather than adjusting it per task. The key diagnostic is understanding why Claude fails: if the model lacks the capability to solve a problem even with full context, upgrade the model; if Claude skips necessary steps or bails on multi-step work, increase the effort level.
Anthropicexplains the technical foundation: model weights (parameters) are fixed during training and determine what the model 'knows,' while effort controls Claude's willingness to invest tokens in exploration, verification, and iterative problem-solving rather than asking the user for more context.
- Higher effort means Claude will read more files, run more tests, and push further through multi-step tasks before checking back in
