Concepts

Reasoning

Some models think before they answer. You can set how hard they think, and read the reasoning they return.

Which models

Check capabilities.reasoning in GET /v1/models. In the launch catalog these support reasoning, in both variants:

  • Kimi K3.1 Pro (utmost/kimi-k3.1-pro, large)
  • DeepSeek V4.1 Flash (utmost/deepseek-v4.1-flash, medium)
  • Qwen3 30B A3B (utmost/qwen3-30b-a3b, small)

reasoning_effort

{
  "model": "utmost/deepseek-v4.1-flash",
  "reasoning_effort": "high",
  "messages": [{ "role": "user", "content": "Prove that the halting problem is undecidable." }]
}

Values: none, minimal, low, medium, high. The value is passed to the backend. What each level does depends on the model.

  • On a model without reasoning support, reasoning_effort is dropped and the request runs normally. There is no error.
  • With an auto alias, reasoning_effort raises the difficulty estimate (medium +1, high +2), so the router leans toward larger tiers. See Routing.

In /v1/responses use reasoning: {"effort": "high"}. In /v1/messages, thinking: {"type": "enabled"} sets effort to high; budget_tokens is ignored.

Reading the reasoning

Chat completions return it in message.reasoning_content, when the model provides it:

{
  "message": {
    "role": "assistant",
    "reasoning_content": "Assume a decider H exists...",
    "content": "Suppose a program H could decide..."
  }
}

When streaming, it arrives in delta.reasoning_content, before the answer's delta.content.

Endpoint Where reasoning appears
Chat completions message.reasoning_content / delta.reasoning_content
Responses A reasoning output item, with the text in summary[0].text. When streaming, it arrives whole (output_item.added then output_item.done), with no deltas.
Messages A thinking block (thinking_delta when streaming). signature is empty.

Billing

Reasoning tokens are output tokens and are billed at the output price. They are included in completion_tokens and reported separately in usage.completion_tokens_details.reasoning_tokens.

Reasoning is not replayed. In Responses, reasoning input items are accepted and dropped. In Messages, thinking blocks in earlier turns are dropped. In chat completions, reasoning_content on earlier assistant messages is passed through to the backend as sent.