Reasoning

Models can produce internal reasoning (step-by-step thinking) before their final answer. Reasoning improves quality on hard problems but adds latency and output tokens, so it is off by default — you get fast, direct responses unless you opt in.

The control is the same on both endpoints and for every model.

Enable or disable per request

Add a thinking object to the request body:

{
  "model": "kimi-k2.7",
  "messages": [{ "role": "user", "content": "..." }],
  "thinking": { "type": "enabled" }
}

Value	Behavior
{ "type": "enabled" }	The model reasons before answering.
{ "type": "disabled" }	No reasoning (the default).
omitted	No reasoning (the default).

This works identically on POST /v1/chat/completions and POST /anthropic/v1/messages.

If you use the OpenAI standard reasoning_effort parameter (low, medium, high), it is also honored as a request to reason.

Reading reasoning back

OpenAI endpoint (/v1/chat/completions) — reasoning is returned in reasoning_content, and its token count in usage.completion_tokens_details.reasoning_tokens:

{
  "choices": [{
    "message": {
      "role": "assistant",
      "reasoning_content": "First, I need to ...",
      "content": "The answer is ..."
    }
  }],
  "usage": {
    "completion_tokens": 320,
    "completion_tokens_details": { "reasoning_tokens": 256 }
  }
}

When streaming, reasoning arrives in delta.reasoning_content before the answer arrives in delta.content.

Anthropic endpoint (/anthropic/v1/messages) — reasoning is returned as a thinking content block, before the text block:

{
  "content": [
    { "type": "thinking", "thinking": "First, I need to ..." },
    { "type": "text", "text": "The answer is ..." }
  ]
}

max_tokens and reasoning

Reasoning tokens count toward your output budget. With reasoning enabled and a small max_tokens, the model can spend the whole budget thinking and return little or no visible answer. When you enable reasoning, set a generous max_tokens.