How do I control reasoning?

Reasoning behavior varies by model. For some models, reasoning is enabled by default; for others, it’s opt-in; and some models may not support turning reasoning off. The supported models list shows each model’s reasoning support, and the reasoning docs list the supported reasoning_effort values for each model, such as none, low, high, or max, depending on the model.

Reasoning tokens are generated separately from the final answer and may appear in reasoning_content. They count as output tokens for billing and toward TPM, so high-reasoning requests can use far more tokens than the visible final answer suggests. If costs, latency, or rate limits are surprising on a reasoning model, reasoning-token usage is a common cause.