One dial controls the cost,
the speed, and the depth.
The effort parameter reached general availability on 5 February 2026 with Opus 4.6, gained an xhigh level in April, and by the time Opus 5 launched it had become the primary way of controlling how hard Claude thinks. Understanding this one setting is the single most useful thing we teach right now.
The ladder runs low, medium, high, xhigh, max. On Opus 5, thinking is on by default and can only be disabled at effort high or below. That sentence quietly retired the on/off mental model most teams still carry.
More effort is not more correct
The instinct is to reach for the top of the ladder whenever a task feels important. In practice a well-specified low or medium request handles straightforward drafting, formatting and lookup just as well, faster and cheaper. xhigh and max earn their cost on genuine ambiguity and multi-step reasoning: complex code changes, careful financial modelling, anything where a quietly wrong answer costs more than a slow one.
The right level depends on the task, not the model
The same team using Claude for quick Slack summaries and client-facing analysis needs to make that distinction consciously, ideally with defaults set per workflow rather than left to individual judgement each time. Unmanaged, teams converge on either always-cheapest, quietly accepting worse output on hard tasks, or always-highest, paying for reasoning nobody needed.
Doing nothing is no longer doing it cheaply
Opus 4.8 defaults effort to high; Opus 5 has thinking on by default. Teams migrating from models where thinking was off by default can see cost and latency rise simply from upgrading, unless effort levels are reviewed as part of the migration rather than inherited. The 5-series migration guide treats this as a first-class step.
Where xhigh fits
xhigh arrived with Opus 4.7 in April, deliberately filling the gap between high and max. Teams that had been defaulting to max out of caution, because high did not quite feel sufficient, are exactly who it is for. If that is you, revisit your defaults; the middle option now exists. More in the Opus 4.7 guide.
Teach the judgement, not the setting
The mechanics take five minutes. The habit that pays is matching level to task: a written default per workflow, an agreed exception rule for when to step up, and a periodic check that the defaults still fit. That is a training exercise, not a configuration one.
List your five most common Claude tasks and the effort level each currently runs at. If you cannot fill in the second column, the setting is choosing for you, and it is probably choosing wrong in at least one direction.
Things people
usually ask.
It has been progressively replaced. Current models manage reasoning adaptively, steered by the effort parameter; on Opus 5, disabling thinking altogether is only permitted at effort high or below. The toggle mental model no longer matches how the models behave.
For most day-to-day business use, medium or high, with a deliberate step up to xhigh for genuinely hard tasks. The wrong answer is any default nobody chose. Set it per workflow, write it down, and revisit it after each model upgrade.
It costs more tokens and time on the same task, yes. Whether it costs more overall depends on what a wrong answer costs you. A financial model checked at max effort can be far cheaper than the same model quietly wrong at low.