which claude effort level should you actually use?
there is a setting in claude's model menu, next to the send button, that decides how hard it works on every reply and how fast you burn through your plan. anthropic published its own numbers on it on september 8, 2026, and they cut against the instinct to turn it all the way up.
claude's effort level sets how much work it puts into each reply: how long it thinks, how much it writes and how many tools it calls. high is the default and the right place to start. drop to medium or low for routine, repeatable work, raise it only when claude cuts a corner, and save max for your hardest problems.
what the effort dial actually changes
effort sets how many tokens claude spends on a reply, and that covers everything it produces: the thinking, the visible answer and every tool call. anthropic's developer docs say it applies to all output tokens, which is why it matters even when thinking is off (Anthropic developer docs, accessed 10 sep 2026). lower effort also means fewer and terser tool calls. it is a behavioural signal rather than a hard budget: on a genuinely hard problem claude still thinks at low effort, just less than it would at high.
you will find it in three places. in the claude apps it sits in the model menu next to the send button, alongside the model and thinking settings (Anthropic help center, accessed 10 sep 2026). in claude code it is /effort. on the api it is one setting on the request. the help center lists it for fable 5.1, opus 5, sonnet 5, fable 5, opus 4.8, opus 4.7, opus 4.6 and sonnet 4.6.
the part most people miss is the cost. anthropic's help center says higher effort gives more thorough responses that take longer and use more tokens, so you reach your usage limits faster (Anthropic help center, accessed 10 sep 2026). on a subscription, effort is how fast your allowance drains. on the api, it is your bill.
the five effort levels, side by side
there are five levels, and high is the default: setting high is exactly the same as not setting effort at all (Anthropic developer docs, accessed 10 sep 2026). the middle column is anthropic's description. the right two columns are how i map each level to work a small business actually does.
| level | what anthropic says it is for | use it for | watch for |
|---|---|---|---|
| low | the most efficient level, with some capability reduction; simpler tasks and subagents | reformatting, sorting, tagging, summarising a document you already trust | can scope a multi-step job too narrowly |
| medium | a balance of speed, cost and performance | everyday drafting, research summaries, edits you can describe precisely | test it on real examples before it becomes a habit |
| high (default) | high capability, the default everywhere | anything new, anything client-facing, anything you have not tested lower | nothing, this is the baseline |
| xhigh | extended capability for long-running agentic and coding work, over 30 minutes | a long claude code or cowork task you will leave running | not on every model; built for token budgets in the millions |
| max | maximum capability with no limit on token spend | the one genuinely hard problem this week | cost, latency, and over-thinking simple work |
xhigh is the one to check before you reach for it. anthropic lists it for fable 5.1, mythos 5.1, fable 5, mythos 5, opus 5, opus 4.8, opus 4.7 and sonnet 5, and notes that not every model with max also has xhigh (Anthropic developer docs, accessed 10 sep 2026).
what anthropic's september 8 numbers actually show
a newer model at low effort can match an older model at high effort for a fraction of the cost, and turning effort all the way up buys less than it costs. both findings come from anthropic's own post on cutting cost without losing performance, written by lance martin (Anthropic, 8 sep 2026).
claude fable 5.1 at low effort matched claude fable 5 at high effort on cursorbench 3.2, at a third of the cost (Anthropic, 8 sep 2026). that is anthropic reporting on its own models on a coding benchmark, not an independent test, so read it as a strong vendor signal rather than a law.
the second result is about the top of the dial. on the 50 hardest frontiercode diamond tasks, claude fable 5 scored 11.5% at low effort, at $5.35 per task (Anthropic, 8 sep 2026). at max effort the same model scored 30.9%, at $19.00 per task (Anthropic, 8 sep 2026). that is roughly 2.7 times the score for roughly 3.5 times the cost (Anthropic, 8 sep 2026).
two things to notice. the model in that second result is fable 5, not 5.1. and those are the hardest tasks in the set, which is exactly where extra effort earns the most. if max buys 2.7 times the score on problems built to be hard, it is difficult to see it paying for itself on a status update. i explained what that benchmark measures in what is frontiercode.
the same post names the failure on the other side directly. high effort can make claude over-think, spending more time deliberating than the task warrants, which adds cost and latency and can make the answer worse (Anthropic, 8 sep 2026).
you are probably already turning effort up by hand
if your saved prompts say "be thorough" or "check your work twice", you are doing the effort dial's job inside the prompt, and anthropic's data says it costs you. the same post lists thoroughness and emphasis boosters, and verification rituals, among the prompt anti-patterns to remove when moving to a newer model (Anthropic, 8 sep 2026).
anthropic tested this on a customer support benchmark during a move from opus 4.8 to opus 5. removing those anti-patterns cut cost by 14.6% on average (Anthropic, 8 sep 2026). it also raised accuracy by 5.3% on average (Anthropic, 8 sep 2026).
the fix is not a better adjective. it is to delete the instruction and set the level. i went through which old instructions to delete first in context engineering for claude 5.
how to pick your level, in five steps
tune effort by type of work, once, rather than deciding it fresh on every task. anthropic's claude code team recommends starting from each model's default and adjusting as a general preference for the kind of work you do, not task by task (Lydia Hallie, Anthropic, 7 jul 2026). their headline guidance is one line:
"for most tasks you should use the model's default effort level" (Lydia Hallie, Anthropic, 7 jul 2026)
- start at the default. high is the default in the claude apps, claude code and the api. leave it there until you have a reason to move it.
- move routine jobs down, then check. take a job you run every week, run three to five real examples at medium or low, and compare them with high. if you cannot tell the difference, keep the lower level. anthropic's help center says low and medium work well for routine tasks and stretch your usage further.
- raise it only for a named failure. anthropic's guidance is to go higher when claude got it wrong by skipping a file, not running the tests, or bailing on a job partway through (Lydia Hallie, Anthropic, 7 jul 2026). the business version: it skipped a section of the brief, missed rows in a list, or stopped halfway.
- re-test after every model upgrade. a setting tuned on an older model carries old assumptions. anthropic's docs tell anyone moving to opus 5 to run a fresh effort sweep rather than reuse old settings, and the fable 5.1 result above is the reason.
- set it at the start of a session. on most models, changing effort mid-session makes the next request re-read the whole conversation with no cache hits (Anthropic Claude Code docs, accessed 10 sep 2026). fable 5.1 is the exception: with an api key or a claude subscription, claude code keeps the cache when you change effort, since version 2.1.260 on 3 sep 2026. why that matters for your bill is in why claude fable 5.1 is cheaper.
the operator move: write the level into the skill
if a job always needs the same effort, put the level in the skill or command that runs it, so nobody has to remember. claude code has supported an effort setting in the frontmatter of skills and slash commands since march 2026, and it overrides the session's level whenever that skill runs (Claude Code changelog v2.1.80, 19 mar 2026). a fix on 9 sep 2026 made that setting apply on opus 4.7, opus 4.8 and fable 5, where it had been ignored (Claude Code changelog v2.1.267, 9 sep 2026). if you set effort in a skill and wondered why nothing changed on those models, update claude code.
two related settings help. /effort remembers a default for each model separately, so switching models does not drag one model's level onto another (Claude Code changelog v2.1.251, 28 aug 2026). and a maxEffortLevel setting, added on 9 sep 2026, caps the highest level anyone can pick in a given setup, while still allowing lower ones (Claude Code changelog v2.1.267, 9 sep 2026).
this is the difference between knowing about a dial and building it into how the work runs. it is the same argument as prompts vs skills: the setting that matters should live in the file, not in your memory.
effort is not the same decision as model choice
the model decides how capable claude is. effort decides how hard that model works on this particular reply. when something goes wrong, the kind of wrong tells you which dial to turn. anthropic's rule of thumb: if claude had all the relevant context, clearly tried, and still got it wrong, pick a more capable model (Lydia Hallie, Anthropic, 7 jul 2026). if it got it wrong by cutting a corner, raise effort instead.
i have written about matching the model to the job in stop using one ai model for everything, and compared the current top two in opus 5 vs fable 5. effort is the cheaper experiment of the two: you change one setting, keep your model and your workflow, and see whether the output holds.
the takeaway
leave effort at high until you have a reason to move it. then move routine, repeatable work down and check it, raise the level only for a failure you can name, and write the level into the skill so it stays put. running everything at max feels safe, and anthropic's own numbers say it can make the answer worse, not better.
common questions
what effort level should i use in claude for everyday business work?
start at high, the default. for routine, repeatable jobs like reformatting, sorting or summarising a document you already trust, run three to five real examples at medium or low and compare them with high. if you cannot tell the difference, keep the lower level. anthropic's help center says low and medium work well for routine tasks and stretch your usage further.
does a higher effort level use up my claude usage limit faster?
yes. anthropic's help center says higher effort produces more thorough responses that take longer and use more tokens, so you reach your usage limits faster. effort applies to every output token, including thinking, the visible answer and tool calls, so running routine work at max spends your allowance on deliberation the task did not need.
is max effort always better than high in claude?
no. max removes the limit on how many tokens claude spends, which helps on genuinely hard problems. anthropic's september 8, 2026 post warns that high effort can make claude over-think, adding cost and latency and sometimes making the answer worse. on the 50 hardest frontiercode diamond tasks, fable 5 at max scored about 2.7 times its low-effort score for about 3.5 times the cost.
should i change claude's effort level in the middle of a conversation?
usually not. pick it at the start. on most models, changing effort mid-session makes the next request re-read the whole conversation with no cache hits, so you pay full price for it again. the exception is claude fable 5.1 with an api key or a claude subscription: since claude code version 2.1.260 on september 3, 2026, changing effort keeps the cache.
want the routine jobs set up once?
the ai builder toolkit is a set of claude skills you install and run, so the jobs you repeat every week run the same way every time. no course, no call.
see the toolkittuning this alongside other people is easier than doing it alone. the ai builders lounge is where the weekly builds happen.
or just follow along. new field notes most weeks on x, instagram, and tiktok.