Opus 5.5 aces all tasks at every effort from low to max

The article grew out of a gap between documentation and practice. Effort is described in API documentation as a way to adjust how many tokens Claude uses in a response, affecting thinking, tool calls and the length of explanations — but the author found that reading the docs alone did not show how the level changes what actually happens on a task, so the tests were run to make that visible.

The effort ladder runs low, medium, high, xhigh and max, and the default differs by model: Opus 5.5 defaults to medium, Opus 4.7 to xhigh, and others such as Fable 5.1 and Sonnet 5 to high. Anthropic's documentation explains why Opus 5.5 sits one step lower — it says Opus 5.5's medium is equal to or better than Opus 5's high on coding and knowledge-work evaluations, and advises starting at medium instead of carrying over the previous model's setting. Settings can be changed mid-conversation from the slider next to the send button, and can also be set per model in settings.json, per session via the CLAUDE_CODE_EFFORT_LEVEL environment variable, or per skill and subagent in front matter. A max setting is not saved as a default and lasts for that session only. Thariq's article adds that changing effort in Claude Code does not break the prompt cache on new models.

The two tasks were deliberately built with traps — a fruit-shop sales CSV set that mixes Shift_JIS files, full-width numbers and stray total rows, and a 24-meeting time-zone conversion that mixes differing daylight-saving switch dates, southern-hemisphere summer time, a +5:45 offset and 12-hour notation. Both were solved correctly at all five levels, three runs each, with the visible difference appearing in self-checking: medium and above reconciled the August total row each time, high and above verified unit prices by dividing amount by quantity, and max reopened the output file. On the time-zone task, max checked its tzdata version, read the update history on the web, and recomputed all 24 rows with GNU date as a cross-check, while xhigh did the first step only once. The author notes two earlier security-leaning tasks were also solved at low, and reads this as consistent with Anthropic lowering the default to medium. The practical reading is that effort decides who does the double-checking — the model or the user — so the setting is likely to matter most for readers who want to step away from a task, while those staying in the loop may find the lower levels sufficient.