Short version: start at the default (medium) for most work, step up to high or xhigh when a task is hard and you can check the result, and save max for problems where a wrong answer costs far more than the tokens.
Effort controls how much Opus 5.5 thinks before it answers. Thinking can't be switched off on this model, so effort is the main dial for quality, speed and cost. Going by Anthropic's docs and third-party tests:
- low: fastest and cheapest. Fine for classification, extraction, quick rewrites and sub-tasks. One test found it clearly weaker on longer deliverables.
- medium (the default): the right starting point for most coding and knowledge work. Anthropic reports it matches or beats Opus 5 at high.
- high: for complex multi-file changes and careful analysis.
- xhigh: for hard agentic coding. Turns get much longer.
- max: uncapped thinking. In one game-building test, max took 22 minutes and $5.25, against 52 seconds and $0.56 at medium.
What you get for the extra spend: in one published comparison of tasks solved out of 100, medium scored 51, high 54, xhigh 56 and max 58, while usage rose from 1x to about 4.5x. Each step up buys less than the one before. These figures come from vendors and small tests, so run a few of your own tasks at two levels before you settle on one.
🕒 As of October 2026
