Claude Effort Levels Explained: Low, Medium, High, xhigh, Max — and When Higher Just Burns Your Limits (2026)

I use Claude Code every day, and "effort" is the setting people fumble most — right after usage limits. It sits one click away in every Claude app and in the terminal, it visibly changes how the model behaves, and almost nobody knows what the levels actually mean. A question that keeps coming up captures it exactly: "How does 'Effort' in Claude models affect the output?" — asked by someone on the free tier running "Sonnet 5 with Max Effort and Thinking," watching a single message drain the limit and wondering whether the lower levels would "butcher" their work.

That thread is full of good instinct and a few myths. So here's what effort actually does, grounded in Anthropic's own docs, with the community's hard-won defaults layered on top — and one popular Reddit belief corrected along the way.

The one-line version

Effort is a dial for how many tokens Claude is willing to spend on your request — across everything it does, not just its private thinking. Higher effort means more thorough reasoning, more tool calls, and longer answers; lower effort means faster, cheaper, terser responses that skip the deliberation on easy work.

Here's the whole ladder at a glance:

high is the default — setting effort to high behaves exactly like not setting it at all. Everything below is a step down for savings; xhigh and max are steps up for the hard stuff.

Effort and Thinking are two different knobs

This is the single most common confusion, and the Reddit thread had it too — people treat "Max Effort and Thinking" as one super-setting. They're separate, and they stack.

• Thinking is whether Claude works through its reasoning in an expandable block before answering. In the apps it's a toggle; you see a timer and a collapsible "Thinking" section.
• Effort is how hard Claude works on the whole response — thinking included, but also the answer itself and any tool calls.

The official help center draws the line cleanly: "the effort level controls how thorough Claude is with every response. The thinking toggle controls whether Claude works through its reasoning" in that visible section. On the current models, higher effort makes Claude think more and more often — at high, xhigh, and max it will almost always think deeply; at low it may skip thinking entirely on a simple problem.

Where to find it

In the Claude apps (claude.ai and the desktop/mobile apps), click the model name next to the send button. The effort menu shows up on the newer models — Opus 4.6 and up, plus Sonnet 4.6 and Sonnet 5 — with a Thinking toggle nested underneath. You can change it mid-conversation. (On an Enterprise plan, if a level is missing, your admin may have turned it off.)

In Claude Code, it's the /effort command — /effort high, /effort max, and so on. The terminal defaults to xhigh, which is why Claude Code feels more thorough (and more expensive) out of the box than a casual chat.

The myth: effort is not a fixed token budget

One commenter in the thread said max is "32k for max as far as I understand" — a fixed thinking budget per level. That's the old mental model, and it's worth un-learning, because it leads to bad decisions.

Effort is a behavioral signal, not a strict token budget. Claude doesn't get handed "you have N tokens to think"; it gets a sense of how eager it should be to spend, and it still scales to the actual problem. At lower effort it will still think hard on a genuinely difficult question — just less than it would at a higher level for the same question. The old fixed-budget approach (budget_tokens) is deprecated or outright removed on every current model; effort replaced it precisely because a per-request budget was the wrong abstraction.

The practical upshot: don't imagine you're "buying 32k tokens of thinking" at max. You're telling Claude don't hold back on this one — and on an easy task, that mostly means it burns tokens you didn't need to spend.

Why higher effort drains your limit faster

This is the part the original poster was feeling in real time, and it's the most important thing to internalize. Effort affects all tokens in the response — text, tool calls, and thinking. Anthropic says so plainly: higher effort "uses more tokens, so you'll reach your usage limits faster."

Because Claude's plans meter tokens, not messages, running everything at max is how you end up rate-limited after a handful of prompts — exactly the "one message drains the limit, then wait 5 hours" experience from the thread. The heavier the model, the sharper this bites: multiple commenters flagged that Sonnet 5 burns through tokens fast at the higher effort settings, and that "high as your default" is the setting that keeps you out of trouble.

When each level is actually worth it

The docs and the people who use this daily land on roughly the same playbook:

• low — simple, scoped, or high-volume work: quick lookups, classification, short edits, and subagents doing narrow jobs. One thread regular put it well: for dropping data into a database or light text updates, low is fine.
• medium — the drop-in for the average workflow when you want good results while spending less. Anthropic pitches it as "comparable to Sonnet 4.6 at high effort" on Sonnet 5, and several users report medium quietly handling most of their real work.
• high — the default and the balance point: complex reasoning, difficult coding, most agentic tasks. If you set nothing, you're here.
• xhigh — the recommended starting point for serious coding and agentic work on the models that support it (Opus 4.7/4.8, Sonnet 5, Fable 5). It's built for long-running tasks over 30 minutes with token budgets in the millions — repeated tool calls, deep search, big refactors. Expect visibly higher token use than high.
• max — reserve for genuinely frontier problems. Anthropic's own guidance is blunt: on most workloads max "adds significant cost for relatively small quality gains," and on some structured tasks it can even overthink. One user's take from the thread — "anything past high on Fable is not worth it… I'm even beginning to think it thinks too much past high and does strange things" — is directionally in line with the docs.

Note that xhigh and max aren't on every model. xhigh is Opus 4.7 and newer (Opus 4.7/4.8, Sonnet 5, Fable 5); max covers those plus Opus 4.6 and Sonnet 4.6. Older models cap out lower. If a level you expect isn't in the menu, the model is usually the reason.

A note on Claude Code's "ultracode"

If you use Claude Code and see ultracode in the effort menu, don't mistake it for a sixth level. It isn't a higher effort setting — it's xhigh effort plus standing permission for Claude Code to spin up multi-agent workflows on its own. Same reasoning depth as xhigh; the extra is orchestration, not intelligence.

FAQ

What's the difference between effort and thinking?

Thinking is whether Claude shows its reasoning in an expandable block before answering. Effort is how thorough the entire response is — thinking, the answer, and tool calls. They're independent toggles and can be combined; effort is the bigger lever on cost.

Which effort level should I use?

Leave it on high (the default) for most real work. Step down to medium or low for simple or high-volume tasks to save your limit. Go up to xhigh for long, tool-heavy coding runs, and max only for genuinely hard problems where correctness matters more than cost.

Does higher effort use more of my usage limit?

Yes — directly. Effort spends more tokens across the whole response, and plans meter tokens, so higher effort means you hit your 5-hour and weekly limits faster. This is the number-one reason to not default everything to max. See our usage limits guide for how the metering works.

Is max effort a fixed token budget?

No. Effort is a behavioral signal, not a strict budget. Claude still scales to the problem — max just tells it not to hold back, which on easy tasks mostly wastes tokens. The old fixed-budget_tokens approach is deprecated on current models.

Why isn't xhigh (or max) showing up for my model?

Level availability depends on the model. xhigh is Opus 4.7 and newer; max is available on Opus 4.6/4.7/4.8, Sonnet 4.6, and Sonnet 5. On older models the menu is shorter. On Enterprise plans, an admin can also disable levels.

Bottom Line

Effort isn't complicated once you stop thinking of it as a token budget and start thinking of it as how hard should Claude try. High is the default and the right answer most of the time. Reach down to medium and low to stretch your limit on routine work; reach up to xhigh and max only when the problem genuinely warrants it — and remember that every notch up spends more of the same pool that runs out after a few big prompts. The person in that Reddit thread had the right fear (a "lesser" setting butchering their work) pointed at the wrong knob: for most tasks, high with thinking on is already the careful choice — max is just the expensive one.