AI Tools Review Editorial TeamQuick answer:
Claude has five effort levels: low, medium, high, xhigh and max. Effort changes how many tokens Claude spends on everything it outputs, including tool calls and thinking. Most models default to high; Opus 5.5 and Haiku 5.5 default to medium. Anthropic recommends starting at the model's documented level, running your own evals, and using xhigh or max only where they demonstrably help. One creator's test found low and max scored the same on 17 hidden tests at very different cost.
The effort dial is the cheapest quality-per-pound lever in Claude, and most people either leave it alone or crank it to max. Neither is what Anthropic's own documentation advises.
Here is exactly what each level does, the default and starting point for every current model, the hard API rules that return errors, and a simple way to choose.
Bart Slodyczka runs the same build at all five effort levels in Claude Code and grades each with hidden tests.
Executive Summary
Effort is the single most useful dial on modern Claude models and one of the least understood. It is not a thinking budget and not a response-length setting. According to Anthropic's documentation it is a behavioural signal that changes how many tokens Claude spends across everything it outputs: explanations, tool calls and their arguments, and thinking when it is active. Get it right and you cut cost and latency with no loss of quality; get it wrong and you either pay for work you did not need or ship shallow answers on hard problems.
The five levels are low, medium, high, xhigh and max. Defaults differ by model, recommended starting points differ again, and some interactions (thinking modes, prompt caching, per-message changes) have hard rules that return errors if you break them. This guide pulls those rules together from Anthropic's effort page, adds the practical test results one creator published this week, and ends with a simple decision process you can apply today. If you are tuning a specific model, pair it with our guide to prompting Claude Opus 5 and our Claude Opus 5.5 review.
What Effort Actually Is
Anthropic describes effort as a way to trade off response thoroughness against token efficiency with a single model. It is set with output_config.effort on a Messages API request, or on the model object for Managed Agents, and in Claude Code with the /effort control. The top-level parameter is available on all supported models with no beta header.
Three points matter most. First, effort applies to all tokens in the response, so it works with or without thinking. Second, it is a behavioural signal, not a strict budget: at low effort Claude still thinks on sufficiently hard problems, just less than it would at higher effort. Third, on newer models such as Opus 5, effort controls thinking volume rather than visible response length, so changing it does not reliably shorten answers; if you want shorter replies, prompt for length.
The Five Levels, Defined
- max: absolute maximum capability with no constraint on token spending. For tasks needing the deepest possible reasoning.
- xhigh: extended capability for long-horizon work: long-running agentic and coding tasks (over 30 minutes) with token budgets in the millions.
- high: spends as many tokens as the task needs for excellent results: complex reasoning, difficult coding, agentic tasks. The default on every supporting model except Opus 5.5 and Haiku 5.5.
- medium: a balanced approach with moderate token savings. The default on Opus 5.5 and Haiku 5.5, for agentic tasks balancing speed, cost and performance.
- low: the most efficient, with significant token savings and some capability reduction. Suited to simpler tasks and subagents.
Support varies. Anthropic's page lists max on Fable 5.1, Mythos 5.1, Fable 5, Mythos 5, Opus 5.5, Opus 5, Opus 4.8, Opus 4.7, Opus 4.6, Sonnet 5.5, Sonnet 5, Sonnet 4.6 and Haiku 5.5, while xhigh is available on a narrower set. The documentation states plainly that not every model that supports max supports xhigh, and that its per-model recommendations override the generic table where they differ.
Defaults and Recommended Starting Points by Model
Anthropic's guidance is explicit that you should run a fresh sweep on your own evals rather than copy settings between models, because levels are recalibrated release to release.
Opus 5.5
Supports all five levels and defaults to medium, one level below Opus 5 and earlier Opus models, so unchanged code now runs lighter. Adaptive thinking is always on and cannot be turned off; a request withthinking: {"type": "disabled"} returns a 400 error at every level. Effort is therefore the primary control over both reasoning and cost. Use a large max_tokens at higher levels because it is a hard limit on thinking plus response text. Opus 5
All five levels, default high. Anthropic suggests stepping up to xhigh for demanding coding and agentic work, max when a task justifies unconstrained spending, and using low and medium liberally as the main cost and latency control wherever evals show quality holds. Thinking cannot be disabled at xhigh or max. For xhigh and max, start around 64kmax_tokens and tune. Sonnet 5.5
All five levels, default high on the Claude API, with recalibrated levels. For agentic coding and multistep tool use, start at medium for well-specified tasks and move to high for harder or longer ones; for chat or latency-sensitive work, start at medium or low. It also offersthinking: {"type": "between_tools"} as the lowest thinking setting, which works at low, medium and high but returns a 400 at xhigh or max. Haiku 5.5
All five levels, default medium. Anthropic says to start at medium even for agentic coding, use low for chat and simple high-volume requests, and warns that in long agent prompts the model is more likely to skip a search, stop early or skip a check at low. Use xhigh or max only where evals show gains, and compare against Sonnet 5.5.Fable 5, Fable 5.1 and Mythos
Start at high, the default; use xhigh for the most capability-sensitive work and step down to medium or low for routine tasks. Anthropic notes lower effort settings on Fable 5 still perform well and often exceed xhigh performance on prior models.Older models
Opus 4.7 and 4.8: start at xhigh for coding and agentic use, high as the minimum for intelligence-sensitive work, medium for cost-sensitive workloads, and max only with measured headroom. Opus 4.7 follows low and medium more strictly than 4.6 and scopes its work to what was asked. If you see shallow reasoning, raise effort instead of prompting around it. Sonnet 4.6 defaults to high but Anthropic recommends explicitly setting medium for most applications to avoid unexpected latency.Effort, Tool Use and Thinking
With tools, lower effort tends to combine operations into fewer calls, make fewer calls, skip preamble and use terse confirmations. Higher effort tends to make more calls, explain plans before acting, produce detailed summaries and write fuller code comments. That matters for agent cost: a verbose, many-call agent at high effort can cost several times a terse one on the same task.
With thinking, the thinking parameter decides whether Claude thinks before answering, while effort controls how much work goes into the whole response, including how often and how deeply it thinks in adaptive mode. At lower levels Claude can skip thinking on simple problems. In a tool loop, follow-up requests that only process tool results can skip thinking at any level. Do not pass adaptive as an effort value; it is a thinking mode, not a level.
What a Real Test Showed
Anthropic's docs tell you to measure on your own workload. Creator Bart Slodyczka did exactly that in a video published on 30/10/2026: he gave Claude Code the same build at all five levels on Opus 5.5 and graded each with 17 hidden tests. By his account, low took 44 seconds and cost 18 cents, while max took 40 minutes and $9.50 (about £7.30), with the same score. His conclusion, headlined as a challenge to common advice, is to start on medium and raise effort only where it demonstrably pays, and he suggests asking one diagnostic question when Claude gets something wrong rather than reflexively turning the dial up.
Treat that as a useful anecdote, not a benchmark: it is one task, one model and one grader, and Slodyczka himself describes where high effort earns its keep. But it lines up with Anthropic's own wording on Opus 4.7 that max adds significant cost for relatively small gains on most workloads and can overthink less intelligence-sensitive tasks.
Prompt Caching and Changing Effort Mid-Conversation
Changing the top-level effort value between requests invalidates the prompt cache. Anthropic therefore advises holding it constant within a conversation that relies on cache hits, and varying it across workloads instead. On Fable 5.1, Mythos 5.1, Opus 5.5, Opus 5, Sonnet 5.5 and Haiku 5.5 you can use a per-message effort change (in beta) that preserves the cache; on Opus 5 it is not available on Amazon Bedrock. There is a catch: with between_tools on Sonnet 5.5, or disabled thinking on Haiku 5.5, a per-message effort that differs from the level in effect returns a 400 error, so use adaptive thinking if you want per-turn variation.
A Practical Decision Process
1. Start at the documented starting point for your model, not at max. 2. Build a small eval of 20 to 50 real tasks with a pass/fail check. 3. Sweep one level down and one up and record pass rate, tokens, latency and cost. 4. Pick the cheapest level that holds quality, and use low for subagents and simple high-volume steps. 5. Reserve xhigh and max for long-horizon agentic runs where evals show a gain, with max_tokens set generously. 6. When an answer is wrong, diagnose first: is it missing context, an ambiguous instruction or genuinely hard reasoning? Only the last justifies more effort. 7. Re-run the sweep on every model upgrade, because levels are recalibrated.
Limitations and Caveats
Effort is not a guarantee: it is a signal and Claude may still think at low effort on hard problems. It does not reliably control output length on Opus 5. Third-party guides disagree on level names and defaults, so rely on Anthropic's page, which also changes as models ship. Cost savings depend on your task mix, and anecdotal tests like the one above should be replicated on your workload. Finally, hard API rules (400 errors for disabling thinking at xhigh or max on certain models) mean effort changes should be tested in staging, not flipped in production.
Choosing Between Effort and Model Size
Effort and model choice are two cost levers. Anthropic positions Haiku 5.5 for high-volume, latency-sensitive classification, extraction and routing, Sonnet 5.5 as the best speed-intelligence mix, and Opus 5.5 for long-running agentic coding and knowledge work. A common pattern is a cheaper model at medium for the bulk of steps and a larger model at high for the hard ones. See how the top models compare in our Claude Opus 5.5 vs GPT-6 Astra comparison and our Claude Opus 5 review.
Who Should Care
Claude Code users running long sessions should check their default and drop to medium for well-specified tasks. API developers should set effort explicitly and sweep it per model. Agent builders should run subagents at low or medium and keep the planner higher. Teams budgeting AI should treat effort as the first lever before model switching. Explore more tooling in our AI tools directory.
The Bottom Line
Higher effort is not automatically better. The documented defaults, recommended starting points and the available test evidence point the same way: start at the model's recommended level (medium on Opus 5.5 and Haiku 5.5, high on most others), measure on your own tasks, and only pay for xhigh or max where the numbers justify it. That is the cheapest quality improvement most teams have not yet made.
Last updated: 31/10/2026. Sourced from Anthropic's effort documentation (platform.claude.com) and Bart Slodyczka's video of 30/10/2026. Pound conversions are approximate; verify the current docs before changing production settings.
Get the free guide: Claude vs ChatGPT, Gemini & Grok
A 20-page playbook covering everything you need to choose and use the big four AI models in 2026, full cost and feature comparisons, what each is best (and worst) at, and how-tos for images, vectors, building a website, Claude Code and more.
