Reasoning models

Reasoning ("thinking") models work through a problem internally before answering, which helps with math, coding and complex analysis. Reasoning tokens are usually billed at the output price.

OpenAI models

Control reasoning depth with reasoning_effort (low, medium, high; some models support more levels):

In the Responses API the field is reasoning.effort.

Claude models

Enable extended thinking in /v1/messages:

Chinese models

Dedicated reasoning models such as deepseek-reasoner need no extra parameters; the reasoning is returned in reasoning_content. Kimi, Zhipu GLM, MiniMax, Qwen and others use their own thinking switches, which the gateway forwards to the upstream unchanged.

Reasoning tokens count toward usage. The dashboard's usage records show the reasoning effort and token counts for each request.