Claude's Effort Setting, Explained

If you typed "effort" into an AI-tools search, you have almost certainly run into Claude's effort control - a setting that tells Claude how hard to think before it answers. It is one of the more useful and least-understood levers in the whole product, because it directly trades answer quality against speed and cost.

According to Anthropic, Claude's thinking is now adaptive: the model evaluates each request and decides for itself whether to reason and how much. The effort setting is the primary lever you use to steer that decision - it sets Claude's overall posture, from "skip deliberation and answer fast" up to "think as deeply as needed." It shows up as an effort menu in the Claude apps and Claude Code, and as an effort parameter in the API.

This guide explains what the effort setting actually controls, how it relates to the separate "thinking" toggle, where you change it, and - most importantly for a business - when a low effort level is the smart, cheaper choice and when a high effort level earns its extra tokens. We reviewed Anthropic's official documentation so the terminology and behavior here match what the product does today, not what a blog post from a year ago claimed.

What is Claude's effort setting?

Claude's effort setting is a control that tells Claude how thoroughly to reason before responding - higher effort means more thorough answers, lower effort means faster, cheaper ones. Anthropic describes effort as "the primary steering lever for thinking": it sets a default for how often Claude thinks and how deeply on each request.

Anthropic's documentation lists five effort levels, from lowest to highest reasoning: low, medium, high (the default), xhigh (extra high), and max. Each level shifts Claude's willingness to spend reasoning on a task. At low, Claude minimizes thinking and skips it for simple tasks where speed matters most; at max, Claude always thinks with no constraints on how deep that reasoning goes.

The key thing to understand is that effort is soft guidance, not a hard dial. Claude still decides per request whether reasoning will help - a simple factual question may get a direct answer even at a higher effort level, while a genuinely hard problem triggers deeper reasoning. Effort raises or lowers the baseline; it does not force a fixed amount of thinking on every message.

• low - Claude minimizes thinking and answers quickly on simple tasks.
• medium - moderate thinking; may skip it for simple queries.
• high (default) - Claude almost always thinks; deep reasoning on complex tasks.
• xhigh - always thinks deeply with extended exploration; built for long-running coding and agentic work (Anthropic notes availability on Opus 4.7 and newer).
• max - always thinks with no cap on depth; the most thorough option.

How does effort differ from the "thinking" setting?

Effort and thinking are two separate settings, and you can combine them freely. Anthropic states plainly that "thinking and effort are separate settings, and you can use any combination of the two." They answer different questions.

The effort level controls how thorough Claude is with every response - how much reasoning it is willing to do under the hood. The thinking toggle, by contrast, controls whether Claude shows you its reasoning in an expandable section before it answers. One governs how hard Claude works; the other governs how much of that work you get to see.

In practice this means you can run high effort with the thinking display turned off (thorough answers, clean output), or low effort with thinking shown (fast answers where you still want to glance at the logic). For most business users the effort level is the setting that changes results; the thinking toggle is mostly about visibility and trust.

How it works: the cost, latency, and quality trade-off

Higher effort produces more thorough responses, but they take longer and consume more tokens - which means higher cost and faster usage-limit consumption. Anthropic is explicit: "Higher effort means more thorough responses, but they take longer and use more tokens, so you'll reach your usage limits faster."

The reason is that reasoning tokens are billed. When Claude thinks, that internal reasoning counts as output tokens and is billed even though you may only see a summary of it. Anthropic warns that the billed output token count does not match the visible token count in the response - you are billed for the full thinking process, not just the text you read. So a max-effort answer to a hard question can cost several times what a low-effort answer to the same question would.

That is the whole trade-off in one sentence: effort buys you reasoning depth, and you pay for it in tokens, money, and wait time. The skill is spending it where depth changes the answer and saving it where it does not.

Where do you set the effort level?

You set effort in the same place you pick the model - in the Claude apps, in Claude Code, and via the API. Anthropic's instructions for the apps are to click the model name beside the send button, select "Effort," and choose your level, and you can change the model, effort level, or thinking setting at any point in a conversation.

In Claude Code, effort and thinking are exposed as adjustable settings for coding sessions, with low and medium suited to routine tasks, high as the default balance, xhigh for long-running coding and agentic tasks, and max for the deepest analysis. Anthropic's help center is the authority on the exact menu wording, which the tools update over time.

In the API, effort is a request parameter. Anthropic's documentation shows it set at output_config.effort (for example, "output_config": { "effort": "medium" }) alongside adaptive thinking - it is not set inside the older thinking object. This lets developers tune effort per request, so a system can run low effort on routine steps and high effort on the hard ones within the same workflow.

• Claude apps - click the model name next to send, choose Effort, pick a level.
• Claude Code - adjust effort (and the separate thinking setting) for the session.
• API - pass effort as a request parameter (Anthropic documents output_config.effort with adaptive thinking); tune it per request.

When to use low vs high effort for business tasks

Use low or medium effort for routine, high-volume, or time-sensitive work, and high or above for tasks where a wrong or shallow answer is expensive. Anthropic frames low and medium as working well for routine tasks that "stretch your usage further," high as the best overall balance, and the top levels as reserved for the hardest reasoning.

The practical test we give clients: ask whether more deliberation would actually change the output. Formatting a list, drafting a routine email, extracting fields from a document, or answering a simple FAQ rarely benefits from deep reasoning - low or medium gets you the same result faster and cheaper. Multi-step analysis, tricky code, financial or legal reasoning, and anything where an error compounds are where higher effort earns its cost.

Because effort is adjustable per conversation and per API request, the smart pattern for a business is not to pick one level forever. It is to default low or medium for the bulk of routine volume and deliberately step up to high or max for the small number of high-stakes tasks - which keeps spend controlled without capping quality where it counts.

• Low / medium - summaries, routine drafting, data extraction, classification, simple Q and A, high-volume batch jobs.
• High (default) - most day-to-day analysis and coding where you want a solid quality-speed balance.
• xhigh - long-running coding sessions and agentic workflows that reason across many steps.
• Max - the highest-stakes reasoning: complex analysis, thorny debugging, decisions where a shallow answer is costly.

Limits and caveats to know

Effort is guidance, not a guarantee, and it interacts with other settings in ways worth understanding before you rely on it. Anthropic notes that Claude decides per request whether to think, so raising effort does not force reasoning on trivial messages, and lowering it does not stop Claude from reasoning on something genuinely hard.

Two practical caveats matter for teams. First, on the API, thinking counts toward max_tokens, so at high effort and above Claude may exhaust the budget - Anthropic advises either raising max_tokens to leave room for reasoning plus the answer, or lowering effort so Claude thinks less. Second, changing the effort level between requests invalidates prompt caching, so Anthropic recommends picking an effort level per conversation and keeping it, then steering individual turns with prompt wording rather than flipping the parameter.

Finally, availability of levels varies by model, and Anthropic's approach has evolved (from manual extended-thinking budgets to adaptive effort). Treat Anthropic's documentation as the source of truth for which levels your model supports and how the parameter is named, and verify behavior on your own workload before deploying it at scale.