API

Reasoning

Use reasoning_effort or Anthropic's thinking with reasoning models through DigitSpace, and how each maps to the other provider format.

View as Markdown

On /v1/chat/completions, set reasoning_effort. On /v1/messages, set thinking. A request with either only goes to offers that support reasoning, and with auto, to a model of at least the standard level.

reply = client.chat.completions.create(
    model="auto",
    reasoning_effort="medium",
    messages=[{"role": "user", "content": "How many weekdays are there in March 2027?"}],
)
print(reply.choices[0].message.content)

How it maps between formats

When an OpenAI-format request goes to a Claude model, reasoning_effort becomes a thinking budget:

`reasoning_effort`	Thinking budget (tokens)
`minimal`	1,024
`low`	2,048
`medium`	8,192
`high`	24,576
`xhigh`	32,768

The gateway also raises max_tokens to at least the budget plus 1,024, and leaves out temperature and top_p, which Anthropic doesn't allow with thinking. On AWS Bedrock the same mapping applies to Claude models only.

When an Anthropic-format request goes to an OpenAI-format model, thinking becomes reasoning_effort: high for a budget over 16,000 tokens, otherwise medium.

Reading the reasoning

On /v1/chat/completions, Claude's thinking comes back as reasoning_content on the message, or in the stream's deltas.
On /v1/messages, an OpenAI-format model's reasoning comes back as a thinking block, except in streams, where it isn't sent.

Reasoning tokens are output tokens: they're billed at the output price.

Checked against the gateway on 31 October 2026. Something wrong or missing? Tell us.