Quick answer
An effort setting is a per-request control for how much hidden reasoning a model does before answering. Low effort is fast and cheap; maximum effort reasons at length, uses many more tokens, and scores higher on hard tasks. The Claude 5.5 family (Opus, Sonnet, and Haiku 5.5) exposes it directly, and in those models thinking cannot be turned off and explicit token budgets return errors, so effort is the lever. Anthropic published the cost of one FrontierCode attempt on Sonnet 5.5 at each level: about $0.76 at low and $12.54 at maximum. OpenAI's reasoning-effort parameter and Gemini's thinking budgets are the equivalents.
Choosing a model used to mean choosing a quality level. Effort settings split that decision: one model, several quality levels, priced by how long you let it think.
What the levels do
- Low: little or no hidden reasoning; best for classification, extraction, formatting, and anything with an obvious answer
- Medium: the sensible default for most chat and coding work
- High: longer reasoning for debugging, planning, and multi-step analysis
- Maximum: benchmark-level thinking at benchmark-level cost; use when being right is worth a large multiple of the price
What it costs
Reasoning tokens are output tokens, so effort multiplies the output bill. Anthropic's own figures for Sonnet 5.5 show a sixteen-fold spread between low and maximum on one coding benchmark. The same request can therefore cost less than a dollar or more than twelve depending on a single parameter, which is why effort belongs in your routing logic, not in a default you never revisit.
How to route
- Start every workflow at low or medium and raise effort only where quality measurably improves
- Pair effort with model choice: Haiku 5.5 at high effort may beat Sonnet 5.5 at low for some tasks, at a fraction of the price
- Log the effort level with every request so cost analysis can see it
- In the 5.5 models, remove thinking budgets, sampling parameters, and disabled-thinking flags from old code; they now return errors
Bottom line
Effort is the cheapest optimisation most teams are not using. Measure your tasks at two or three levels, route accordingly, and you will usually cut cost without touching quality.
