Docs / Features

Reasoning tokens

Reasoning models think before they answer; those thoughts are billable output tokens.

Controlling effort

Use the reasoning parameter with an effort level (minimal, low, medium, high, xhigh) where the model supports it. Higher effort improves hard-task accuracy and increases cost and latency.

JSON

{
  "model": "deepseek/deepseek-v4-pro",
  "messages": [{"role": "user", "content": "Prove the theorem"}],
  "reasoning": { "effort": "high" }
}

Reasoning content is returned separately from the final answer where the provider exposes it, and is always billed as output tokens.