Reasoning

A reasoning model writes out its reasoning before it answers. These models list reasoning in capabilities on GET /v1/models. Reasoning tokens are billed as output tokens, and usage counts them separately as reasoning_tokens.

Chat Completions

The reasoning comes back beside content under two names, reasoning_content and reasoning, which always hold the same text. Engines and client libraries disagree on the name, so every model returns both. Read whichever one your client expects.

"message": {
  "role": "assistant",
  "content": "391",
  "reasoning_content": "17 × 23 = 17 × 20 + 17 × 3 = 340 + 51 = 391.",
  "reasoning": "17 × 23 = 17 × 20 + 17 × 3 = 340 + 51 = 391."
}

When streaming, the reasoning arrives first as deltas carrying both fields, then the answer as delta.content:

data: {"choices":[{"index":0,"delta":{"reasoning_content":"17 × 23","reasoning":"17 × 23"},"finish_reason":null}],...}
data: {"choices":[{"index":0,"delta":{"content":"391"},"finish_reason":null}],...}

stream = client.chat.completions.create(
    model="moonshotai/kimi-k3",
    messages=[{"role": "user", "content": "What is 17 × 23?"}],
    stream=True,
)
for chunk in stream:
    if not chunk.choices:
        continue
    delta = chunk.choices[0].delta
    reasoning = getattr(delta, "reasoning_content", None)
    if reasoning:
        print(reasoning, end="")
    if delta.content:
        print(delta.content, end="")

Note

A client that shows only delta.content shows nothing while the model reasons. That is the reasoning step, not a stalled stream. max_tokens covers the reasoning and the answer together, so a low limit can end the reply inside the reasoning, with finish_reason: "length" and no content.

You don’t need to send the reasoning back in later turns. If an assistant turn you send fills both fields, reasoning_content is the one used.

Responses

The reasoning is an output item of type reasoning, ahead of the message. Its text is in content[].text (type reasoning_text) when the model returns it; summary[].text holds a summary instead, when one is produced. Usage counts it in output_tokens_details.reasoning_tokens.

Messages

Ask for the reasoning with thinking: { "type": "enabled" }, or "adaptive", which is accepted identically. It comes back as a thinking content block ahead of the text; when streaming, it arrives in content_block_delta events whose delta.type is thinking_delta. Without it you get the answer alone, unless the model put its whole reply inside the reasoning; that turn comes back as an ordinary text block. See Messages.

Reasoning effort

Send a level in reasoning_effort on Chat Completions, reasoning.effort on Responses, and reasoning_effort or output_config.effort on Messages. Each model lists the levels it takes in reasoning_efforts on GET /v1/models. The applied level comes back in X-Reasoning-Effort. How each endpoint validates the level is described on its own page.