Free tools Windows power users keep installed
One-click scans. No signup required.
Start with medium. It is Claude Opus 5.5’s default and the best baseline for most coding, analysis and knowledge-work requests. Move down to low for simple, speed- or cost-sensitive work; move up to high, xhigh or max only when your own evaluation shows that deeper reasoning improves the result enough to justify extra tokens and latency.
Opus 5.5 has five effort levels—low, medium, high, xhigh and max. Effort changes how much reasoning, tool use and self-checking Claude applies. Adaptive thinking remains enabled at every level, so effort—not a separate thinking switch—is the main control over reasoning depth and request cost.
As an Amazon Associate I earn from qualifying purchases.
What the effort setting changes
Effort is a per-request control for the amount of work Claude is allowed to do before returning an answer. At lower settings, Opus 5.5 generally reaches a result with fewer reasoning tokens, fewer tool calls and lower latency. At higher settings, it can spend more tokens exploring alternatives, checking intermediate results, using tools and repairing mistakes. The trade-off is that a harder-working response can cost more and take longer.
This is not a simple accuracy slider. A high-effort response can be unnecessary for a well-specified one-line transformation, while a low-effort response can be a poor fit for a repository-wide code change or a multi-step investigation. Effort also interacts with your prompt, available tools, context, token cap and success criteria.
#1 Best Overall
Adaptive thinking is always on for Opus 5.5, including at low. The thinking control cannot be disabled. Selecting an effort level therefore sets the practical budget and depth for that adaptive process rather than turning reasoning on or off.
Opus 5.5 levels at a glance
| Level | What it is for | Typical trade-off | Good starting examples |
|---|---|---|---|
low |
Simple tasks, rapid iteration and cost-sensitive workloads | Lowest token use and latency, with some capability reduction on difficult tasks | Formatting, extraction, short summaries, straightforward edits and quick prototypes |
medium |
General-purpose work; the Opus 5.5 default | Balanced quality, cost and response time | Routine coding, analysis, documentation, research synthesis and ordinary agent turns |
high |
Difficult coding, analysis and agentic work | More thorough reasoning, with higher token use and latency | Debugging interacting components, planning migrations and checking complex logic |
xhigh |
Long-running or unusually difficult coding and agent tasks | Extended reasoning can be expensive and slow; gains are workload-dependent | Tasks expected to run for more than 30 minutes, large refactors and difficult investigations |
max |
The hardest tasks where maximum capability matters | Deepest analysis and no effort-side constraint on token spending; highest latency and cost exposure | Critical, ambiguous or failure-expensive work after a benchmark shows a real quality gain |
How to use each level
low: efficient first drafts and short feedback loops
Choose low when the task is narrow, the answer is easy to verify, or you expect many quick iterations. It is useful for converting data, applying a known code pattern, classifying items with clear labels, generating test scaffolding or making small documentation edits.
Anthropic reports that low comes close to Claude Opus 5 at high on several coding evaluations while using much less effort. That finding is specific to the published evaluations, not a guarantee that every coding task will behave the same way. If your low-effort run starts missing edge cases or requires repeated correction, measure whether medium is cheaper overall once rework is included.
The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →medium: the default baseline
medium is the recommended starting point for Opus 5.5. Anthropic says it matches or exceeds Opus 5 at high on its coding and knowledge-work evaluations. Use it as the control condition in your own tests before deciding that another level is better.
Medium is usually appropriate when the task matters but is not obviously adversarial: implementing a feature with tests, explaining an unfamiliar module, drafting a technical plan, comparing design options or running a normal tool-using workflow. It gives you a meaningful quality baseline without immediately paying for maximum reasoning.
high: more checking for difficult work
Select high when a wrong answer is costly, the problem has several interacting constraints or the model must inspect and revise its own work. Examples include debugging a failure that crosses services, designing a data migration, tracing a security-sensitive behavior or producing an analysis that must reconcile conflicting evidence.
Expect more output tokens, tool calls and elapsed time than at medium. The setting is worthwhile only if those extra steps improve your measured success rate, reduce human review or prevent expensive rework.
xhigh: extended agentic tasks
xhigh is intended for long-running coding and agentic work, particularly tasks lasting more than 30 minutes. It gives the model room to sustain a larger plan, inspect more artifacts and revisit earlier decisions. Reserve it for cases where your evaluations show a quality improvement; simply making a short request xhigh does not make the prompt better.
Because long runs can consume substantial budgets, define a stopping condition and a way to verify the final state. For code, that may mean a required test command and a clean diff. For an investigation, it may mean a list of claims with supporting evidence and explicit unresolved questions.
max: maximum capability for exceptional cases
max is the deepest Opus 5.5 setting. It is appropriate when the task is unusually hard or failure is more expensive than additional compute—for example, a complex architectural decision, a large autonomous refactor or a critical analysis with many dependencies.
Max does not impose an effort-side constraint on token spending, but your account limits, rate limits and any max_tokens value still apply. Verify the quality gain before making max your default. If the same tests pass at high, paying for max is difficult to justify.
Choose effort by task, not by prestige
| Work pattern | Start here | Escalate when |
|---|---|---|
| Deterministic transformation or short answer | low |
Output needs repeated correction or misses stated constraints |
| Routine coding or knowledge work | medium |
Tests fail, important edge cases are missed or the plan is incomplete |
| Multi-file debugging, migration or tool workflow | high |
High still fails representative cases and the task can tolerate more time |
| Agent run expected to exceed 30 minutes | xhigh |
Only after xhigh demonstrates a measurable gain should you consider max |
| Failure-expensive, highly ambiguous work | high or xhigh |
Use max when an evaluation shows its extra quality is worth the cost |
A useful escalation rule is to increase effort after observing a specific failure mode, not because the task “sounds advanced.” Keep the prompt and tools stable, raise one level, and check whether the failure disappears. If it does not, the problem may be missing context, an unclear success criterion or a tool limitation rather than insufficient effort.
Benchmark the levels on your own workload
There is no universal low-to-max accuracy or cost percentage that applies to every user. Anthropic’s published comparisons are based on its own coding and knowledge-work evaluations. Results vary with task design, prompt, tools, context, token cap and scoring method, so an effort sweep on representative work is more informative than a generic ranking.
- Build a fixed task set. Select real prompts that represent your workload, including easy, typical and failure-prone cases. Keep the original requirements and reference inputs unchanged.
- Hold the environment constant. Use the same model version, system instructions, tools, context, success criteria and temperature or other sampling controls for every run. Change only the effort level.
- Run all five levels. Test
low,medium,high,xhighandmaxwhere your platform exposes them. Repeat tasks when variability matters instead of treating one lucky response as a result. - Set a sufficient token cap. The
max_tokenslimit covers thinking tokens as well as visible response text. A cap that is too low can truncate high-effort reasoning and make a setting look worse than it is. Anthropic’s prompting guidance says that 128,000—the model maximum cited in its 2026 documentation—worked well for long agentic coding turns. - Capture operational metrics. Record task success, factual or test errors, latency, total and output tokens, number of tool calls, retries and actual cost. Include human review time when that is part of your workflow.
- Score against explicit criteria. For code, use automated tests, linting, type checks and review findings. For analysis, use a rubric for correctness, completeness and unsupported claims. Mark a run failed when it violates a hard requirement even if the prose looks polished.
- Choose the lowest level that meets the target. Compare total work, including retries and review. A faster first response is not cheaper if it creates manual repair; a max response is not better if it adds no measurable value.
Keep a small regression set after choosing a default. Re-run it when prompts, tools or model versions change so an effort decision remains evidence-based rather than anecdotal.
Token limits, context and latency
Opus 5.5 has a 1-million-token context window according to Anthropic’s 2026 model documentation. Context capacity is separate from reasoning effort: a request can contain a very large codebase or document set without requiring max, and a short but logically difficult prompt can benefit from high effort.
At high, xhigh and max, leave enough output budget for both internal thinking and the visible answer. If the model stops early, returns an incomplete patch or fails to finish a tool loop, inspect the token cap before concluding that the effort level is incapable. Raising the cap can help, but it also increases the maximum amount the request may consume.
Higher effort generally increases latency because Claude may reason longer and call tools more often. For interactive products, measure p50 and tail latency rather than relying on a single response time. A medium default with selective escalation often gives a better user experience than running every request at max.
Prompt caching and changing effort
Changing top-level effort between requests invalidates the prompt cache. If a long shared context is cached, that invalidation can add work and cost on the next request. Anthropic’s prompting guidance notes that changing effort per message can preserve the cache for later turns. Check how your client maps effort to requests before designing a benchmark, and record cache behavior as part of cost and latency measurements.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What Anthropic’s published tests establish
Anthropic reports two useful directional results: medium Opus 5.5 matches or exceeds Opus 5 at high on coding and knowledge-work evaluations, and low comes close to Opus 5 at high on several coding evaluations at lower cost. These statements compare specific model-and-effort combinations on Anthropic’s evaluations; they do not establish a universal ranking of all five Opus 5.5 levels.
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Clear out junk files and repair common Windows errors3Scan for outdated or missing drivers - takes under a minuteNo independent, general low-to-max accuracy or cost figure can be stated safely. Do not treat a claimed percentage, token multiplier or speed ratio as portable unless it comes from a benchmark that matches your prompts, tools, context and accounting method.
Best Value
Practical operating policies
For an interactive coding assistant
Use medium for normal edits and explanations. Escalate a turn to high when tests expose a cross-file issue or the requested change has several constraints. Use xhigh for a planned, long-running refactor with a clear verification command. Keep max for incidents or designs where your evaluation proves that high or xhigh is insufficient.
For batch processing
Start with low for well-defined records and sample the results against a human-checked set. Move to medium when quality misses exceed your threshold. Reserve high or above for ambiguous records, and route only those cases instead of paying the maximum effort for the entire batch.
For autonomous agents
Define tool permissions, a time budget, a completion test and a rollback path before increasing effort. More reasoning cannot compensate for a tool that returns incomplete data or permissions that prevent the required action. Log every tool call and intermediate failure so you can distinguish an effort problem from an environment problem.
Troubleshooting common surprises
- Low is fast but unreliable: Move to medium, then compare failure rates on the same cases. Improve the prompt if requirements were ambiguous.
- High appears to do nothing extra: Your task may be too easy to benefit. Check whether tool calls, tokens or rubric scores actually changed before paying more.
- Xhigh or max times out: Increase the permitted token and wall-clock budgets only if the task justifies them; otherwise step down and split the work into verifiable stages.
- The response is truncated: Raise
max_tokensand remember that thinking consumes part of that limit. - Costs jump after an effort change: Check prompt-cache invalidation and whether a top-level setting change forced the shared context to be processed again.
- Results vary between runs: Repeat the benchmark, fix sampling and tool conditions, and report a distribution rather than one best or worst output.
Bottom line
Set Opus 5.5 to medium first. Use low when speed and economy dominate, high for difficult reasoning, xhigh for long-running agentic work and max only when a controlled evaluation shows that its extra effort prevents failures worth more than the added cost and delay. The right setting is the lowest one that reliably meets your own success criteria.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
