Skip to main content

Reasoning

Many models can think before answering. More reasoning usually improves quality on hard tasks, at the cost of speed and budget.

Per request​

On Chat Completions and Responses, send reasoning_effort:

{
  "model": "openai/gpt-5.6-sol",
  "reasoning_effort": "high",
  "messages": [{ "role": "user", "content": "..." }]
}

You can also send a reasoning object, for example { "reasoning": { "effort": "low" } }. Standard Compute passes these settings to the model.

Some models always reason. If you try to turn reasoning off for one of them, the gateway uses a low effort instead of failing the request.

Dashboard defaults​

You can set a default reasoning level per model under Customize on the Models page. Options depend on the model. The default applies only when the request doesn't include reasoning_effort or reasoning.

Messages API​

The Anthropic thinking parameter is not supported on the Messages API. It is ignored, and responses don't include thinking blocks.