The effort parameter directly controls how much computation Claude Opus 4.5 applies to a task. It accepts "low", "medium", and "high" values. Higher effort means the model allocates more internal reasoning and output tokens to a problem, improving depth and precision at the cost of additional time and tokens. Lower effort reduces reasoning steps and response length, giving faster, simpler answers. Opus 4.5 is the only model in the Claude family that exposes this adjustable-effort mechanism.
Benchmark results show concrete effects of effort. With medium effort, Opus 4.5 matches the top performance of Sonnet 4.5 while using far fewer output tokens due to better token efficiency. High effort yields stronger reasoning scores and more detailed planning, which is useful for debugging, large code changes, and complex decision-making. Low effort is ideal for lightweight tasks like rewriting text, extracting fields, or producing short summaries. Because effort influences both token usage and latency, it becomes a key dial for production tuning.
Effort also affects how Claude interacts with tools. At low effort, the model may skip unnecessary tooling steps and produce shorter reasoning traces. At high effort, Claude may issue multiple retrieval calls, compare alternatives, and generate multi-step plans. In agent workflows involving retrieval—such as searching a vector database like Milvus or Zilliz Cloud—higher effort gives Opus 4.5 more space to integrate retrieved information and form a better solution. You can tune effort per endpoint or per task type, allowing more thoughtful computation where needed while keeping interactive tasks fast and economical.