Anthropic shipped Claude Opus 4.8 on May 28, 2026, and tucked into the launch — next to the benchmark charts everyone screenshots — is a small control that changes how you’ll actually use the thing day to day. It’s a dial called effort, and for the first time it’s sitting right there in the claude.ai sidebar next to the model picker, not buried in an API parameter only developers touch.
The launch posts all name the levels. Almost none of them tell you which one to pick when. So here’s the part that’s missing: a plain-English guide to the effort dial, what each setting actually trades away, and the three or four cases where cranking it to “Max” is just lighting tokens on fire for no extra quality.
What the effort control actually is
Think of effort as a “how hard should you think about this?” slider. On higher settings, Claude thinks more often and more deeply before it answers — better results, slower, more tokens. On lower settings it answers faster and sips your rate limits more slowly, at the cost of some depth.
Anthropic puts it plainly in the launch post: “On higher effort settings, Claude will think more frequently and more deeply to give better responses. On lower effort settings, Claude will respond faster and use up a user’s rate limits more slowly.” The control is available on all plans — Free, Pro, Max, Team, Enterprise.
Here’s the one thing that trips people up. The names are different depending on where you are:
- In claude.ai and Cowork (the chat app), the control sits next to the model selector. The default is high. You can bump it up to “Extra” or “Max.”
- In Claude Code and the API, the same dial has five named levels:
low,medium,high,xhigh,max. The thing claude.ai calls “Extra” is exactlyxhighin Claude Code.
So it’s one control with two vocabularies. Anthropic’s effort documentation is the source of truth: “By default, Claude uses high effort, spending as many tokens as needed for excellent results. You can raise the effort level to max for the absolute highest capability, or lower it to be more conservative with token usage.”
The five levels, and what each one is for
This is the table the launch coverage skipped. Each description below is Anthropic’s own wording from the effort docs, paired with the kind of task it’s built for.
- Low — “Most efficient. Significant token savings with some capability reduction.” Use it for simple classification, quick fact lookups, or high-volume jobs where a marginally better answer isn’t worth the extra latency or spend. If you’re running the same prompt across 5,000 rows, this is your setting.
- Medium — “Balanced approach with moderate token savings.” Solid performance without the full cost of high. Good for routine drafting, summaries, and everyday questions where you want it quick but not careless.
- High (the default) — “High capability. Equivalent to not setting the parameter.” This is where Opus 4.8 sits unless you change it, on every surface including the API and Claude Code. It’s tuned for complex reasoning, nuanced analysis, and difficult coding — the stuff most people actually open Claude for.
- Xhigh / Extra — “Extended capability for long-horizon work.” Built for advanced coding and agentic work that needs extended exploration: repeated tool calls, deep searches, the kind of task that runs for a while on its own. Anthropic recommends “extra” specifically for “difficult tasks and long-running asynchronous workflows.”
- Max — “Absolute maximum capability with no constraints on token spending.” The ceiling. No budget, no shortcuts — it spends whatever it takes.
Anthropic’s official recommendation, copied straight from the Opus 4.8 docs: “Start with xhigh for coding and agentic use cases, use high for most other intelligence-sensitive workloads, and step down to medium or low only when you’ve measured that the lower level holds quality on your evals.”
Translation for the rest of us: coding agent → xhigh; everyday thinking → leave it on high; bulk or trivial work → drop to medium or low only after you’ve checked the cheaper setting still gets it right.
A few worked examples
The dial makes more sense when you map it to a real task:
- “Summarize these three meeting notes into action items.” → Low or Medium. It’s a known, bounded job. High effort won’t make the action items meaningfully better; it’ll just take longer.
- “Draft a client proposal section, then tighten the tone.” → High (the default). This is the sweet spot the default is designed for.
- “Refactor this 600-line module, keep the tests green, explain every change.” → Xhigh / Extra. Hard coding with exploration — exactly what xhigh exists for.
- “Plan and run a migration across my repo overnight while I sleep.” → Xhigh / Extra, occasionally Max. Long-horizon, asynchronous, no human in the loop to catch a lazy answer.
- “What’s the capital of Australia?” → Low. Please don’t spend Max effort on Canberra.
One real note from the field: developer @syqrel posted that Opus 4.8 on high effort actually used less of his Pro plan limit than Opus 4.7 did — a reminder that “higher model version” and “higher effort” aren’t the same lever. You can be on the newest model and still dial effort down.
What this means for you
If you’re a regular claude.ai user (writer, marketer, analyst): You almost never need to touch this. Leave it on High. The only time to move it is when you’re doing something repetitive and want speed — then drop to Medium. “Extra” and “Max” are for long, hard, agentic tasks; on a normal chat they mostly just make you wait.
If you use Claude Code: Set /effort xhigh for real coding sessions and agent runs — that’s Anthropic’s own recommendation. Use /effort high (the default) for quick edits and questions. Reserve max for the genuinely gnarly stuff, because it will spend tokens accordingly.
If you run high-volume API jobs: Start at medium or even low, then run your evals. Anthropic’s guidance is explicit that you should only step down once you’ve measured that the cheaper level holds quality. Don’t assume; test on a sample, compare, then roll out.
If you’re on a tight rate limit (Pro plan): Effort is now your throttle. Lower effort = your limit lasts longer. If you keep hitting the cap mid-afternoon, dropping non-critical chats to Medium buys you more sessions.
If you manage a team: This is a training moment. The single most useful sentence to put in your internal doc is: “Leave it on High. Use Xhigh for coding agents. Only go higher when you’ve got a reason.” Most wasted spend on day one comes from people maxing the dial “just in case.”
What the effort dial can’t do
Worth being honest about the limits, because the launch hype skips these:
- It’s not a quality guarantee. Max effort makes Claude try harder, not be right. On a question it would get wrong at High, it can still get wrong at Max — just more expensively and slower.
- Higher isn’t free. Xhigh and Max trade speed and tokens for depth by design. One tester warned that running the highest settings on dynamic, looping tasks produces a “staggering” token burn. Great for a hard refactor; wasteful on a pet project.
- The claude.ai labels aren’t fully documented. Anthropic standardizes the five names (
low–max) in the docs and Claude Code, but hasn’t published an exact slider-label table for the consumer UI beyond “extra” and “max.” If your app shows a “Faster ↔ Smarter” style control, that’s the same dial — just a friendlier face on it. - It won’t fix a vague prompt. A bad instruction at Max effort gets you a thorough, expensive answer to the wrong question. Effort amplifies your prompt; it doesn’t replace it. (That’s what prompt engineering is still for.)
- It doesn’t change the model’s knowledge or context window. Same Opus 4.8, same 1M-token context, same training cutoff at every effort level. You’re tuning how hard it works, not what it knows.
The bottom line
The effort dial is the most useful small change in Opus 4.8 for everyday users, precisely because it’s not a benchmark number — it’s a knob you’ll actually turn. The whole guide fits on a sticky note: leave it on High; go Xhigh/Extra for coding agents and long jobs; drop to Medium or Low for bulk or trivial work; only reach for Max when you have a real reason and don’t mind the bill.
If you want to get genuinely fluent with Opus 4.8 — effort levels, when to fan work out to agents, how to keep a long session on track — our Claude Code Mastery course walks through the workflow end to end, and Prompt Engineering covers the half of the equation the effort dial can’t fix: writing the instruction worth spending effort on.
