Techniques & Methods
Effort Setting in plain English.
Also known as: effort level,reasoning effort,adaptive thinking,thinking budget
The one-sentence version
An API control that tells a model how much reasoning to spend on a request, trading speed and cost against capability without changing models.
An effort setting is a per-request dial for how hard a model thinks. At low effort it answers quickly with little or no hidden reasoning; at high or maximum effort it reasons at length, uses more tokens, and scores higher on hard benchmarks. Anthropic's Claude 5.5 family (Opus, Sonnet, and Haiku 5.5) exposes it directly and published the cost of one coding-benchmark attempt at each level for Sonnet 5.5: about $0.76 at low and $12.54 at maximum. OpenAI's reasoning models have an equivalent reasoning-effort parameter, and Google's Gemini uses thinking budgets. The setting replaced an older pattern: in the 5.5 models thinking cannot be switched off and explicit token budgets return errors, so effort is the lever. Practical use is routing by task: low for classification and extraction, medium as the default, high for debugging and planning, maximum only when a benchmark-level answer is worth a benchmark-level bill. Because the same model serves every level, you can tune cost without re-evaluating quality across vendors.