Understand effort levels in Claude Code

Link copied

Attach video time

Effort is a bit more than just thinking. It also controls how much work Claude does to handle your request, including the number of files it reads, how much it verifies its own work, and the number of steps it takes before it returns to you with a response.

The types of effort levels

There are five levels, from lowest to highest: low, medium, high, xhigh, and max. Every model that supports effort starts at high by default, and that's a solid default to begin with. It's tuned per model to the point where most tasks finish without overspending any tokens.

Not every model offers all five. If for some reason you ask for a level a model doesn't support, Claude Code just runs the highest level it does support at or below the level you picked.

The max option works a bit differently than the others, and actually removes any internal constraints on how much the model can spend and reason on tasks, so it'll typically spend way more time doing things before producing an output, and will eat up tons of tokens. Only use this for really hard asks.

The last option, ultracode, isn't actually sixth effort level. It's a Claude Code setting that sends the xhigh effort to the model, and additionally has Claude orchestrate dynamic workflows for very complex tasks. That's worth knowing, but is beyond the scope of this lesson.

How effort levels work

One thing to keep in mind is that this scale is calibrated on a model-by-model basis. high on Sonnet and high on Opus don't necessarily mean each one will do the same amount of work. So whenever you switch models, it's a good idea to start from the default effort level again, and adjust from there.

Higher effort levels cause Claude to be more thorough in its research phase. It will read more files, verify more of its own work, and push further through a multi-step task before it returns a result to you.

Tokens are all billed at exactly the same rate. Thinking, tool calls, and the text Claude writes to you all come out of the same loop at the same price. Higher effort levels cost more because they produce more tokens, not because the tokens themselves are more expensive.

Keep in mind that none of this is deterministic. The same prompt at the same level won't read the exact same files every time, so treat what you see more as of the shape of the behavior around the response and output.

Models vs. effort

Before you decide to change either the model or the effort, check your context first. Was your prompt too vague, or is Claude pointed at the right files, tools, and skills? If you find yourself raising the effort level on a task that shouldn't need it, you may need a more intelligent model instead.

Once your context is solid, the question to ask is whether Claude either didn't know enough or didn't try hard enough.

If Claude Code has all of the context for what it's trying to solve, but still got something wrong, that's a sign that you're probably using the wrong model. But if Claude Code got something wrong because it skipped loading a specific file or didn't run specific tests or ended halfway through a refactor, that's a signal that you need to have a higher level of effort. The opposite is also true; if you find it reasoning for too long and getting the same level of output that a lower effort level could have provided, you may want to downgrade that effort level.

It is also possible to use smaller models at a higher effort level. For tasks that aren't super complex, this can get you nearly the same result at a lower cost instead of using a more intelligent (and expensive) model. You'll need to balance the tradeoffs of speed vs. quality of response. It's a good idea to experiment with different configurations to see how they work and respond to your asks.