Claude Code effort levels: medium vs high, xhigh, max and ultrathink

What Claude Code uses if you set nothing

Claude Code picks an effort level per model. On Opus 5.5 and Sonnet 5.5 — the models its opus and sonnet aliases resolve to on the Anthropic API — the default is medium. On Opus 4.7 it is xhigh. On most other models that support effort, it is high. That means a session on today's default model runs one level lower than sessions on the previous Opus did, which is worth knowing before you compare the two by feel.

How to change it

Inside a session, /effort opens a slider; /effort high sets a level directly, and /effort auto clears your saved choice so the model's default applies again. In the slider, Enter saves the level as your default for future sessions, while s applies it to the current session only. You can also adjust effort with the left and right arrow keys in the /model menu. Outside a session:

claude --effort high                  # one launch
export CLAUDE_CODE_EFFORT_LEVEL=high  # every launch from this shell

For a persistent default, set effortLevel in your settings, or a per-model level under modelSettings. When several are set, an explicit choice (the environment variable, the flag or /effort) wins over settings, and settings win over the model's default.

Which levels you get depends on the model, and the model on the provider

Fable 5.1 and Fable 5, the Opus models from Opus 4.7 onward, and Sonnet 5 and Sonnet 5.5 offer every level. Opus 4.6 and Sonnet 4.6 stop short of xhigh. The model aliases also resolve differently by provider: opus means Opus 5.5 on the Anthropic API, Claude Platform on AWS, Amazon Bedrock and Google Cloud's Agent Platform, but sonnet means Sonnet 5.5 only on the Anthropic API — it resolves to Sonnet 4.6 on Claude Platform on AWS, and to Sonnet 4.5, which has no effort levels at all, on Bedrock and Google Cloud. So the same sonnet session can default to medium on one provider, high on another, and ignore effort entirely on a third. Check the model /model reports before comparing sessions across providers.

ultrathink, ultracode and max are three different things

These get confused constantly. ultrathink is a word you put in a prompt to ask for deeper reasoning on that one turn; it doesn't change your session's effort level, and phrases like "think hard" are just ordinary text to Claude Code. max is the top effort level — the deepest reasoning and the highest token spend — and unless you set it through the environment variable, Claude Code applies it to the current session only. The documentation notes it can show diminishing returns and overthink simple tasks. ultracode is not an effort level at all: it is a setting that has Claude orchestrate multi-agent workflows for substantive tasks, at whatever effort the session runs at. /effort ultracode turns it on; launching with --effort ultracode sets xhigh and turns it on together.

Medium or high?

Claude Code's own guidance describes medium as suited to day-to-day engineering with a clear scope and high to work where verification matters or edge cases are likely, with low for quick exchanges where you review each result. On the newest models the labels also moved: Anthropic reports Opus 5.5 at medium outperforming Opus 5 at high on coding work, and Sonnet 5.5's levels were recalibrated from Sonnet 5's. So the useful comparison is not between level names across models, but between two levels on the model you actually run — on a task you can check.

Where going lower or higher pays

Drop below the default for narrow, low-ambiguity work: a repetitive edit across many similar files, a lookup with one right answer, a subagent handed a scoped investigation. Raise it for genuinely hard problems — a refactor with real architectural trade-offs, a bug whose cause isn't visible in its symptoms. A task ambiguous enough to justify more effort is usually worth running through plan mode as well.

The cost side

Effort compounds across every step of an agentic session, so it is one of the largest levers on what a session costs — often more than the choice of model. See the effort & thinking cost estimator for the cost curve, and effort level reference for what each level is for across the API.