A 12-minute video by Ben AI (Ben van Sprundel) on Claude's effort levels, built on 2 articles from Anthropic's Claude Code team. The thesis: effort is not a quality dial. The model sets how smart Claude is. Effort sets how much time Claude spends reading, testing and checking, and how many decisions it makes without you. Most people should use low far more, and max almost never.
What effort does
The levels are low, medium, high, extra high (xhigh) and max. Intelligence comes from the model you pick (Opus, Sonnet or Haiku). Effort decides how much work Claude puts into the task.
Thariq Shihipar of the Claude Code team compares it to a deadline. Give someone 12 hours for a task and they assume you want them to try very hard on their own. Give them 1 hour for the same task and they assume you want a decent first version to iterate on.
At higher effort, Claude reads more, tests more and checks its own work more before it answers. Lydia Hallie, also on the Claude Code team, adds the part most people miss: higher effort also means more decisions made without you. At low effort, Claude comes back and asks. So effort is also a dial for how involved you want to be:
- Low: Claude keeps coming back to you. You stay in the loop.
- High and above: Claude goes off and decides for you. Good when you're away, less good when you care about the choices.
2 rules from Anthropic frame the rest:
- A bad prompt can't be fixed with more effort. Describe the task well first.
- Switching effort mid-chat is free on the newest models: Claude doesn't reread the whole conversation at full price. Build on high, tweak on low.
The 4 levels
Low: in the loop, fast
Anthropic's rule: low is "for when I want quick responses that are in the loop". Use it in 3 situations:
- You plan to iterate. You react to a first version faster and steer sooner. Ben now uses low to prep his videos.
- The task is simple. In one of Anthropic's tests, the same simple redesign took 1 minute on low and 28 minutes on max, with a similar result.
- You polish work a higher level made. Build the report or dashboard on high, then fix layout, copy and structure on low.
Typical tasks: brainstorming ("15 subject lines for this newsletter"), quick rewrites ("make this email friendlier and half as long"), sorting and tagging (support tickets by issue type), short summaries, rough first drafts.
The catch: low does exactly what you ask and nothing more. For a task with several parts, give Claude a short checklist.
Medium: the default for most work
Thariq uses medium for most of his regular work. Anthropic's guide says to use the model's default for most tasks, and on Opus 5.5 the default is medium. Use it when you can define what a good result looks like and you'll review it once at the end.
Typical tasks: everyday writing (a follow-up email from call notes), repurposing (a video into a LinkedIn post), meeting summaries with decisions and next steps, standard reports in a fixed format, quick research before a call.
If a medium result is weak, Lydia's advice is not to raise the effort first. Check the prompt and the context. Move up only when Claude skipped a step or didn't check its work.
High: when a mistake would cost you
Thariq's rule: high is "for work where verification is important or there are edge cases". Use it when:
- A mistake costs money, trust or a decision made on wrong numbers.
- The task has many edge cases, such as contract terms or the figures in a financial report.
- You verify work done at a lower level. This is how Thariq works: build and iterate on low, then switch to high at the end to check everything.
Anthropic's data-analysis test shows why. On low, Claude tried 1 approach and got the answer wrong every time. On high, it tried 2 approaches, saw that they disagreed, worked out why, and got it right 4 times out of 5.
Typical tasks: reconciling numbers across sources (a revenue report against Stripe), open-ended analysis ("why did churn go up last quarter?"), risk reviews of contracts, fact-checking, long documents such as a 50-page report, and building or fixing skills, because a mistake in a skill repeats on every run.
Length alone isn't a reason for high. In Anthropic's results, straightforward rule-following work gains the least from more effort. What matters is how much judgement and checking the task needs.
Extra high and max: walk away
Thariq uses max "only if I want zero input". At these levels Claude can work for hours on one task. Use them only when all 3 are true:
- You want to walk away and not answer questions.
- The job is long and hard.
- The work is critical, so quality matters more than time and usage limits.
If any one is false, use high. At these levels Claude assumes it must work everything out alone, so it makes many decisions for you. The prompt matters even more: a vague task can run for hours and still give a poor result.
Typical tasks: long unattended research (a year of support calls and tickets turned into churn reasons, an ideal customer profile and a roadmap), correctness-critical reviews (a 50-page agreement before signing), and Claude's "best one-shot" on a high-stakes deliverable such as an investor deck.
Anthropic's docs warn that on most workloads max adds significant cost for small quality gains, and on some tasks it overthinks and the result gets worse. Extra high is the cheaper version: start there, and keep max for the truly critical.
How to pick, in 4 questions
Ask in this order and stop at the first yes:
| # | Question | If yes |
|---|---|---|
| 1 | Will you go back and forth on it, or is it a simple task? | Low |
| 2 | Would a hidden mistake cost money, a client or a wrong decision? | High |
| 3 | Are you handing off a long, hard, critical job and walking away? | Extra high or max |
| 4 | Anything else, or you're unsure | Medium |
Key takeaways
- Effort is time and autonomy, not intelligence. Pick the model for smarts, the effort for how long Claude works and how much it decides alone.
- Low is underused. For iteration and simple tasks it is much faster with a similar result.
- Medium is the default on Opus 5.5. A weak result usually means a weak prompt, not too little effort.
- Pay for high only where checking matters, and often only at the end: build low, verify high.
- Max is rare. All 3 conditions must hold, and it can make results worse.
Notable quotes
"Effort isn't a quality dial. It's a dial for time and how involved you want to be." - Ben AI
"[Max is for] only if I want zero input." - Thariq Shihipar, Anthropic, as quoted in the video
Why it matters for me
I run most of my NicAI work in Claude Code, so the effort setting applies to almost everything I do there. The split the video suggests: low for the back-and-forth on drafts and notes, medium as the base, high for the final check of anything with numbers in it (pipeline figures, invoices, proposals), and max almost never. The "build low, verify high" pattern fits how I already review drafts. The wider Opus 5.5 guidance is in Claude Opus Best Practices.
The video also promotes a free "Effort Optimizer" mod that suggests an effort level for each prompt. I haven't tried it.
Further Reading
