Claude Code has an effort setting, and it is one of the most useful dials you have. Turning it up does not make Claude "smarter" in a vacuum. Anthropic frames it as adaptive reasoning: the model already adjusts how hard it thinks to the task in front of it, and the effort level sets the ceiling on how much reasoning it is allowed to spend.
You set it with /effort inside a session. It sticks for the session, so you can dial it up for a hard refactor and back down for quick edits.
What does each effort level change?
| Effort | What changes |
|---|---|
| Low | Fastest, cheapest. More direct answers, less exploration, fewer tool calls. |
| Medium | Balanced speed vs thoroughness. |
| High (default) | Deep reasoning, more planning, more comprehensive work. |
| XHigh / Extra | Much larger reasoning budget, aimed at difficult coding and long-running tasks. |
| Max | Essentially "use as much thinking as needed". Highest cost and latency. |
What do higher and lower effort do for coding?
The abstract levels become concrete once you watch how the agent behaves on a real task.
Higher effort tends to make Claude Code:
- Investigate more files before editing
- Explore multiple hypotheses before choosing one
- Create more detailed implementation plans
- Run more verification steps
- Use tools more frequently
- Produce more thorough explanations and summaries
Lower effort tends to make it:
- Jump to implementation faster
- Make fewer tool calls
- Do less validation
- Give shorter responses
- Spend fewer tokens overall
Does it improve coding quality?
Usually, yes, but mostly on harder tasks. For simple requests the difference is small; for hard ones it can be large.
| Task | Effort makes... |
|---|---|
| "Rename this variable" | little difference. The work is obvious either way. |
| "Add a button to this component" | little difference. |
| "Refactor this auth system across 20 files" | a real difference. More effort means more context gathered and more alternatives weighed. |
| "Investigate why these tests intermittently fail" | a real difference. The model spends time reproducing and reasoning instead of guessing. |
Anthropic specifically recommends high or XHigh for intelligence-sensitive work, the kind where a missed edge case or a shallow read costs you an hour.
A subtle but important point
Effort is not exactly the same as chain-of-thought length. Anthropic's docs say the setting affects all token spending, including:
- reasoning
- tool calls
- explanations
- function arguments
A higher effort setting makes the agent behave more deliberately overall: it reads more, checks more, and explains more. You see the effect as better engineering judgment rather than longer answers.Common gotchaDo not reach for Max to rescue a vague request.If the result was wrong because the prompt or context was thin, more reasoning just builds a fancier wrong answer. Fix the instructions first, then raise effort.
Which effort level should I use?
Match the dial to the task, not to your mood.
| Task | Reach for | Why |
|---|---|---|
| Quick edits, boilerplate, routine changes | Low | No exploration needed. You want the edit, fast. |
| Everyday feature work in code you know | Medium | A little planning, no deep dig. |
| Serious engineering work | High | The good default. Claude maps the code and weighs approaches. |
| Large refactors, debugging, architecture | XHigh | Long-running work where a missed case is expensive. |
| A hard problem you are stuck on | Max | Throw the full budget at it, and accept that returns diminish and it can spend a lot of time exploring. |
Which effort level fits my project size?
Project size is a proxy for how much context Claude has to hold and how easy it is to break something far away.
- A small script or single file. Low to Medium is almost always enough.
- A mid-size app. Medium for routine work, High when a change touches shared code.
- A large codebase or monorepo. Default to High, go XHigh or Max for anything that crosses module boundaries.
What is the trade-off with higher effort?
Every step up the dial costs more tokens and more wall-clock time, because the model spends more of both before it acts.
The honest framing is not "always use Max for best results". It is: spend effort where a mistake is expensive or the path is unclear, and save it where the task is obvious. If a Medium answer is right, a Max answer is the same answer that cost five times as much.Common gotchaThe classic mistake is pinning the dial at Max and forgetting it, then watching a one-line edit burn minutes and a pile of tokens.Because the setting sticks for the whole session, a single hard task you bumped up for keeps taxing every trivial follow-up. Drop it back to Low or Medium the moment the hard part is done.
One data point worth knowing
This setting is not cosmetic. In 2026 Anthropic acknowledged that changing Claude Code's default reasoning level had contributed to a perceived drop in coding quality. The takeaway: the effort dial has a real, measurable effect on day-to-day results, so it is worth being deliberate about where you set it.
Practical habits
- Start lower than you think, step up if needed. If a Medium pass misses, re-run on High. You will calibrate fast.
- Bump it up before a plan, not after. Set High or XHigh before asking for a big change so the planning happens at the right depth.
- Drop it back down for the cleanup. The follow-up edits rarely need the same budget.
Effort pairs with the rest of this phase: higher effort spends more tokens (see Token Usage & Cost) and leans harder on a well-shaped context (see Context Engineering).