01Claude effort levels explained: what effort is
Claude effort levels explained simply: effort is how much compute you want Claude to spend on a task. Claude always attempts the task, and effort changes how hard it works on it.
Anthropic recently published how its own team uses effort settings. In this episode, Julian Goldie covers what the levels are, what each is for and one real job where the effort should change.
Effort changes how much Claude thinks, how many files it reads, how much it checks its own work and how many steps it takes before it checks in with you. Anthropic compares it to how long you will wait for a worker. Give them 12 hours and they try very hard.
Give them one hour and they deliver the best version they can.
Effort also changes how much Claude does alone. At lower effort it asks you for more context. At higher effort it makes more choices itself.
02The old way: using max for everything
The old way is to open Claude Code and set effort to high, or switch to max every session. Fixing a typo, brainstorming a title or building a whole app all get max. It feels safe, because more effort should mean better results.
It is not quite true. Each step up in effort uses more tokens and takes more time. Anthropic's API docs say that for most workloads max adds a lot of cost for fairly small quality gains, and on some tasks it can lead to overthinking.
The new way is to match effort to the job. Use low for quick work, and go higher when you are checking work that really matters.
- Typo fix: low is enough
- Brainstorming: low
- Checking important work: higher
03The five Claude effort levels explained
There are five levels: low, medium, high, extra high and max. Not every model supports extra high, so you might only see four.
- Low: quick, interactive work such as brainstorming, sketching and easy changes. You check each answer as it comes back.
- Medium: most everyday software work, like building a new feature. In Claude Code, Opus 5.5, Sonnet 5.5 and Haiku 5.5 all default to medium.
- High: jobs where checking matters or edge cases are likely, like fixing bugs or editing a big codebase.
- Extra high: long jobs over 30 minutes that use millions of tokens.
- Max: fully autonomous work on hard problems, like building and checking a whole app end to end. Anthropic calls it Claude's best one shot.
Julian gives business examples. Low is brainstorming content. Medium is adding a new scene type to video skills or a new step to the SEO article pipeline.
04What Anthropic's tests show
Anthropic built a simple fitness app from a short prompt on Opus 5.5. Low took about one and a half minutes and gave a simple workout log and graph. Max took about 67 minutes and built more, like a heat chart, but took over 40 times longer.
On Terminal Bench 3.0, a coding benchmark, hard tasks with Fable 5.1 went from 34 percent at low to 75 percent at the highest effort. In total, passes went from 140 at low to 214 at max out of 370 attempts. That is about 50 percent more, but it used about three times the tokens.
One data analysis task shows the difference. At low effort, Claude picked one method and reported the result. At high, it tried two methods, saw the results change and worked out why before choosing.
At low it got it right zero times in five tries. At high it got it right four times.
05Tips and mistakes with Claude effort levels
Julian shares several warnings from the episode.
- The Claude Code docs say max can show diminishing returns and is prone to overthinking, so test it before using it everywhere.
- In Claude Code, max resets when you start a new session, so you need to set it so it sticks.
- Higher effort helps with missed edge cases but does not fix a wrong overall approach. If Claude heads the wrong way, more effort will not save it.
- Fix your prompt and context first. Anthropic found that once a spec was detailed, results across effort levels were much more alike.
Julian also notes that an article from the Claude Code team says the default is right for most tasks, and effort should be treated as an override. Keeping it simple is fine.
06A real job: add a goals panel to Agent OS
Julian walks through adding a goals panel to the Agent OS. It is a dashboard panel that shows your goals and how far along each one is. Here is where the effort changes.
- Fix the brief first. Get Claude to interview you: where do goals come from, what should the panel show, and where does it sit?
- Build on low. You steer, ask for changes like putting progress bars on top or making the card smaller, and loops come back fast.
- Switch to high once it looks right. Ask Claude to test the panel: does it read goals right, what happens with no goals, do the other Hermes tabs still work?
- Use extra high only when you leave a big job running, like upgrading the whole Agent OS overnight.
The shape is low while you steer, high when it is checking, and the big levels when you walk away. This follows a workflow from the Claude Code team: interview, build on low, verify on high.
07Who needs Claude effort levels explained
This is for anyone using Claude Code who has been leaving effort on high or max out of habit. It saves tokens and time, and it can give better results by avoiding overthinking.
It also helps anyone building with agents. For Julian's team, extra high means rebuilding a full pipeline with many steps and files while nobody is watching. Max means building and testing a brand new agentic operating system end to end, and only after testing it first.
If you want help choosing a model or effort level for your own tasks, Julian says the AI Profit Boardroom has weekly coaching calls where you can bring your setup. The main rule is the one he ends on: match the effort to the job.