Set the level
The reasoning field expresses a preference for extra deliberation. It does not turn every model into a reasoning model or guarantee internal reasoning in the response.
Choose effort for the task: extracting fields often benefits more from a strict schema, while a difficult comparison may benefit from deeper processing.
Reasoning preference
json
{
"model": "openai/gpt-5.6-sol",
"messages": [{"role": "user", "content": "Compare two migration plans."}],
"reasoning": {"effort": "high"}
}Account for constraints
- Check supported_parameters for the selected model.
- Leave sufficient output room for a useful final answer.
- Measure latency and usage separately from ordinary requests.
Evaluate quality
Compare variants on a fixed set of real tasks, not one impressive example.
Do not show a user technical intermediate data as an explanation when it was not intended for output.