Claude’s effort setting controls how much work Claude does on each turn: how much it thinks, how many files it reads and how much it double-checks. It does not make the model smarter. Anthropic’s team suggests low for brainstorming, medium for most feature work, high for bugs and verification, and max only for hard, fully hands-off problems.

Key takeaways

Effort means how thorough Claude is, not how smart; model choice decides knowledge
Low for sketching, medium for building, high for verifying, max for autonomous hard work
Switch mid-session with /effort in Claude Code; on newer models it keeps the prompt cache
If Claude didn’t know enough, change the model; if it didn’t try hard enough, raise effort

What effort actually controls

Think of cleaning a home: a quick tidy, a regular clean or a deep clean. The person and their skill stay the same; only the thoroughness changes. Anthropic’s documentation says effort affects all tokens in a response, including thinking, tool calls and the text itself. It calls effort “a behavioral signal, not a strict token budget.” Levels are low, medium, high, xhigh and max. Defaults vary: Opus 5.5 and Haiku 5.5 default to medium, while most other recent models default to high.

One task, four settings

Take a London software team that needs to fix a checkout bug and add a discount-code feature.

Low: sketch. “Give me three ways to add discount codes.” Quick answers, and you stay in control.
Medium: build. Implement the chosen option from a clear spec. Thariq Shihipar of Anthropic recommends medium for most regular software engineering.
High: verify. Hunt the checkout bug and test edge cases. In Thariq’s own small test on a difficult crash bug, low effort solved it in 0 of 5 runs, while xhigh solved it in 4 of 5.
Max: hands-off. Use only when Claude should work alone on a hard problem, end to end.

How to switch

Claude Code: type /effort and pick a level. On newer models such as Opus 5.5, switching mid-session keeps your prompt cache (Claude Code v2.1.280 or later).
Claude API: set output_config.effort in the request.
Other tools: OpenAI’s reasoning models offer a similar reasoning-effort setting, so the same habit applies.

The diagnostic question

Lydia Hallie from the Claude Code team suggests a simple check when output falls short. Did Claude lack knowledge, or did it lack effort? If it misread an unfamiliar domain, use a larger model. If it skipped a file or didn’t run the tests, raise effort.

Frequently asked questions

Does max effort make Claude smarter?

No. Effort controls how thorough Claude is, not what it knows. For more knowledge, choose a larger model.

Which effort level should I use by default?

Start with the model’s default and adjust per task: low for drafts, high for verification.

How do I change effort in Claude Code?

Type /effort and pick a level. You can switch mid-session.

Does higher effort cost more?

Usually, yes, because it generates more tokens. Anthropic’s blog shows one high-effort example using roughly 7x more tokens.

Sources: Anthropic docs, “Effort” (platform.claude.com); Claude blog, “Choosing a Claude model and effort level in Claude Code” by Lydia Hallie; Thariq Shihipar, “Spending your effort”. [