Claude Code has an effort setting that decides how much work the model does on a task. It is easy to treat it as a "thinking time" dial, but Anthropic's July 7, 2026 guidance describes something broader.

The short advice: start with the default for your model, then change effort as a general preference for the kind of work you do, not for every individual task.

What Effort Actually Controls

Effort affects how long Claude thinks, but also how many files it reads, how many tools it calls, how many checks it runs, and how many steps it takes before it checks in with you. Higher effort generally means more reading, verification and persistence. Lower effort leans toward asking you for more context sooner.

Claude Academy describes five levels: low, medium, high, xhigh and max. The default depends on the model. For example, Anthropic's effort documentation says Claude Sonnet 5 defaults to high effort in Claude Code and in the Claude API.

Two clarifications matter:

  • Effort changes how much work Claude does, not its underlying knowledge or capabilities.
  • Effort shapes token use, but it is not a hard token limit.

In Claude Code you can choose a level by running the /effort command.

Matching Effort to the Kind of Work

Anthropic's recommendation is to treat effort as a standing preference. If your day is mostly small, well-scoped edits, a lower setting may fit. If you spend it on cross-file refactors or debugging unfamiliar code, a higher one probably fits better.

When lower effort fits

Lower effort means fewer tool calls, tokens, turns and checks. It suits tasks where the answer is close at hand and the cost of a miss is small:

  • Renaming a variable in a single file.
  • Explaining a short function.
  • Quick edits where you will review the diff yourself.

Because lower effort favors asking for context sooner, you may see more questions back. That is often a good trade when you want to stay closely involved.

When higher effort fits

Higher effort pays off when thoroughness matters more than speed:

  • Changes that touch many files and need consistent updates.
  • Bug hunts where the cause is not in the file you first opened.
  • Work where tests must be found, run and satisfied.
  • Unfamiliar libraries, where Claude may need to investigate rules before acting.

Anthropic's concrete signal is behavioral: raise effort when Claude skips files, tests or checks that it should have looked at.

Reading the Signs and Knowing the Trade-offs

Both directions have costs. Claude Academy notes that effort set too low can make Claude stop before finishing complex work. Set too high, it can mean more tokens, longer response times, overthinking, or flagging work beyond what you asked for.

A useful diagnostic is to separate an effort shortfall from a knowledge or capability gap:

  • Claude skipped files, tests or checks: raise effort.
  • Claude has the relevant context and still gets it wrong: consider a more capable model rather than more effort.
  • Claude keeps expanding scope or taking too long: try a lower level.

Note that more effort can help with unfamiliar library rules because Claude will investigate them, but that may require substantial reading, and therefore more time and tokens.

Defaults can also change. Anthropic has published an engineering postmortem with a section on a change to Claude Code's default reasoning effort, so it is worth checking what your current default is rather than assuming it stayed the same. The /effort command is the quickest way to confirm and adjust it.

Conclusion

There is no single best effort level. The default is a reasonable starting point, and the right adjustment depends on whether your work rewards speed and tight control or thoroughness and persistence.

Watch what Claude does rather than guessing. If it skips checks, raise effort; if it is slow or wandering, lower it; if it has the context and still errs, the fix is likely a stronger model, not a higher setting.