InferenceCapabilities
Reasoning
How models that reason before replying work, and how to control the depth of reasoning.
Some models reason before they reply: first they work through the task, then
they give the answer. Such models are marked in the catalog and in the response
of the GET /v1/models method with the
"reasoning": true flag. The depth of reasoning is set with the
reasoning_effort parameter, and the same parameter can turn reasoning off.
Parameter
reasoning_effortstring
Reasoning depth:
none,low,mediumorhigh. The valuenoneturns reasoning off completely.
Request
{
"model": "yenisei-max",
"messages": [{"role": "user", "content": "Calculate and explain the steps of the solution"}],
"reasoning_effort": "high"
}
Where to find the reasoning text
The reasoning text comes in the choices[].message.reasoning field, separately
from the reply itself in content.
Credit usage
Reasoning is also made of reply tokens, and it is included in the bill. If you
do not need reasoning, pass reasoning_effort: "none".