Claude Code's Effort Levels, Low to Ultracode: The Cheat Sheet for Which Level Fits Each Task
Claude Code has five effort levels and an ultracode switch, and the one you pick decides how fast you get an answer, how hard Claude checks its own work, and how soon you hit your usage limit. What each level really does, from Anthropic's own docs and tests, and a one-page cheat sheet for which one to use for each job.
Type /effort in Claude Code and you get a slider: low, medium, high, xhigh, max, and an Ultracode toggle under it. Most people never touch it, and the ones who do usually crank it to max and leave it there. Both waste something: time on easy jobs, or quality on hard ones, and a lot of your usage limit either way.
Anthropic has now published exactly what each level changes, with test results. Here is what effort actually does, and the cheat sheet for picking the right level for the job in front of you.
What effort actually changes
Effort doesn't swap the model. It tells Claude how much work you want it to put in: how long to think, how much to test, and how many decisions to make on its own before it comes back to you. Anthropic's engineer Thariq Shihipar ran the same jobs at every level and put it like this: higher effort means more verification, more edge-case testing, and more of Claude's own judgement.
His clearest example: he asked Claude to "build a personal fitness and workout tracker app" at each level.
| Level | Time | What he got |
|---|---|---|
| Low | 1.5 min | A log and a simple graph: a base to build on |
| Medium | 4 min | More to it, more detail |
| High | 11 min | More complex, with more choices made for him |
| Max | 67 min | The most complete version, down to a heat chart |
None of these is wrong. Low gives you a quick starting point while you stay in charge. Max gives you Claude's best attempt in one go, but it makes more assumptions for you and takes 45 times as long.
The cheat sheet: which level for which job
| The job | Level | Why |
|---|---|---|
| Brainstorming, a first sketch, a rename or other small change | low | Quick answers you check yourself. Anthropic's docs list exactly these |
| The first version of a feature you'll review before trusting | low | A fast starting point; check you agree with it, then verify on high (the loop below) |
| Everyday feature work with a clear scope | medium | The default on Opus 5.5 and Sonnet 5.5, and enough for most day-to-day work |
| Fixing a bug in an existing codebase, anything with edge cases | high | For work where checking matters: Claude tests more edge cases and verifies before it answers |
| Code where one missed case is a disaster: sanitizers, parsers, security, performance | xhigh | On an HTML sanitizer test, Fable 5.1 passed 1 of 5 attempts at low and 5 of 5 at xhigh |
| Hard jobs you want done without you watching: build and verify end to end, hunt security holes | max | Anthropic's own example. Test it first: the docs say it can overthink |
| Big audits, migrations, research you want cross-checked | ultracode (on top of any level) | Dozens of Claudes split the job and check each other's work. It uses your limit much faster |
| One hard question in the middle of a session | type ultrathink in the prompt | Deeper thinking on that one turn, without changing your level |
Free, for your email
Get the full setup steps + the copy-paste setup prompt
The rest of this guide (4 more sections, 2 copy-paste prompts) opens right here. The setup prompt, one paste into Claude that sets it all up, comes by email only. Use your real address.
Unlocks: How to switch levels · The loop Anthropic's engineer uses for new features · When ultracode is worth it · The honest caveats
50,000+ follow where these guides come from. No spam, reply “unsubscribe” any time.