Claude Code effort levels tested: the same Opus 5.5 build at low, medium, high, extra high and max

What effort changes

• Model versus effort. The speaker, relaying an Anthropic Claude Code engineer's article, says the model is which "brain" answers and effort is how much work that brain does on your task: how many files it reads, how much it checks its own work, and how far it pushes before coming back to you.
• Higher effort also decides more for you. In Anthropic's fitness-tracker test (low: about 1.5 minutes, a log and a simple graph; medium: 4 minutes; high: 11; max: 67), the max build added a heat chart nobody asked for. Fine if you walk away for an hour, wasted if you will rewrite half of it anyway.
• A spec narrows the gap. When the same engineer had Claude interview him, wrote a detailed spec and rebuilt at every level, the designs and builds came out alike. Once you have made the decisions, effort has fewer left to make.

The speaker's own test

Prompt: build a trip expense splitter, "like a mini Splitwise", on Opus 5.5 in a fresh empty folder each time, graded by 17 hidden tests. One trap: settle the group with as few payments as possible.

• Low: 44 seconds, about 18 cents at API prices. It wrote the whole app in one go and never opened it in a browser.
• Medium: 2.5 minutes, roughly 60 cents. It opened its own app and tested it first.
• High: 2 minutes 41 seconds, roughly 60 cents.
• Extra high: almost 11 minutes, about $25. It checked its maths against about 4,300 random trips and added an undo button.
• Max: about 40 minutes, about $95. About 25 minutes of planning before the first file, then its own tests, plus dark mode and backups nobody asked for.

Every level, including low, passed all 17 tests, including the tricky settlement maths. The speaker's verdict for a clear build: start on low and fix as you go; the extra effort bought polish and self-checking, not a better answer.

When higher effort does pay (Anthropic's tests, as relayed)

• Hidden edge cases. On a Terminal-Bench 3.0 protein-data task with five tries per level, Opus 5.5 on low picked one sensible way to prepare the data and got it wrong all five times; on high it prepared the data two ways, noticed the answers differed, dug into why, and got it right four times out of five.
• Security-style work. A Fable 5.1 task to strip every way of hiding code in a web page passed 1 of 5 at low (about 2 minutes, one test page) and 5 of 5 on a higher setting (about 33 minutes, an attacker-style self-review, a standard security suite and a junk-page generator).
• Limits. More effort cut failures from missed edge cases but not from picking the wrong approach. The biggest gains were in security and hardware; rule-following office work such as a month-end filing gained least.

The decision rule

1. If Claude gives a bad result, look at what you gave it first: a vague prompt, or missing files, tools or skills.
2. If the context was good, ask: did it not know enough, or not try hard enough? If it skipped a file or did not run the tests, raise effort.
3. If it had everything, clearly tried, and was still confidently wrong, that is a knowledge problem. More effort will not fix it; move to a bigger model.

Checked against the Claude Code docs

From Model configuration, read 2026-10-30 (the first part of a long page).

• Levels. low, medium, high, xhigh and max for Fable 5.1 and 5, Opus 5.5, Sonnet 5.5, Haiku 5.5, Opus 5, Sonnet 5, Opus 4.8 and 4.7. Opus 4.6 and Sonnet 4.6 have no xhigh.
• Default. high on every model that supports effort, except Opus 5.5, Sonnet 5.5 and Haiku 5.5, which default to medium, and Opus 4.7, which defaults to xhigh. This matches the video's point that a carried-over "high" habit may cost more than needed on Opus 5.5. A top-level effortLevel in your user settings does not count for Opus 5.5.
• How to set it. /effort opens a slider, /effort <level> sets it, /effort auto clears your saved level; --effort sets it for one session; CLAUDE_CODE_EFFORT_LEVEL sets it by environment variable; effortLevel in settings accepts low, medium, high or xhigh, and max is not accepted there.
• Ultracode. The docs say ultracode is a Claude Code setting that turns on dynamic workflows, not an effort level.
• Not checked: the video's statement that Anthropic says Opus 5.5 on medium matches or beats Opus 5 on high, and the details of the two Anthropic articles.

Caveats

• Costs are API-equivalent estimates from one run per level; the speaker's spoken figures for medium and high are approximate.
• The trip splitter is a clear, small task with one tricky rule. A vague or large task may behave differently.
• The Anthropic examples are toy or benchmark tasks, as the speaker himself notes.