Effort is the single most useful dial on Claude Opus 5, and the one most people leave at its default. It controls how deeply the model reasons and how much it spends getting to an answer — which makes it your primary lever on both quality and cost.
The five levels
| Level | Use it for |
|---|---|
low | Short, scoped tasks; latency-sensitive work that isn't intelligence-sensitive |
medium | Cost-conscious production traffic where quality still matters |
high | The default. Most intelligence-sensitive work |
xhigh | Coding and agentic work — the recommended starting point there |
max | When correctness matters more than cost, and latency doesn't bind |
The parameter lives inside output_config, not at the top level of the request.
The counter-intuitive part
Higher effort does not always cost more overall. On agentic work, a model that reasons harder up front often takes fewer turns, makes fewer wrong tool calls, and needs less correction — so the total token spend for the finished task can be lower at xhigh than at medium.
The relationship is not monotonic, which is exactly why the recommendation is to sweep rather than to assume. Run your own evaluation set at medium, high, and xhigh and look at cost per completed task, not cost per request.
Where Opus 5 differs from earlier models
Two things are worth knowing if you are carrying settings over:
lowandmediumare unusually strong on this model — strong enough that effort defaults inherited from a previous generation are probably wrong. Sweep downward as well as upward.- Effort does not reliably control verbosity. If responses are too long, lowering effort is the wrong lever; a conciseness instruction or an explicit word count is the right one.
One constraint to remember
Disabling thinking is only permitted at effort high or below. A request combining thinking: disabled with xhigh or max returns a 400, and the check applies per request — so a later call that raises effort while thinking is still disabled fails even though earlier calls in the same conversation succeeded.
More on Claude Opus 5
Start with the complete Claude Opus 5 guide for the overview, or go deeper:
Ready to go deeper?
Read the full Claude Opus 5 guideFrequently Asked Questions
What is the default effort level on Claude Opus 5?
`high`, which is also what you get by omitting the parameter entirely. The recommended starting point for coding and agentic work is `xhigh`, and for most other intelligence-sensitive work a minimum of `high`.
Does lower effort make Claude Opus 5 cheaper?
Per request, yes. Per completed task, not necessarily — on agentic work higher effort often means fewer turns and less rework, so total spend can fall as effort rises. Measure cost per finished task rather than per request.
Can I use effort to make responses shorter?
No, not reliably. Effort changes reasoning depth rather than output length. To shorten responses, use an explicit conciseness instruction or a word count — prompting is the effective lever, and a short instruction can cut length by around a fifth.
