Reasoning

Reasoning models think through a problem before they answer. The thinking takes extra tokens, which you pay for, and usually gives better answers on maths, code and multi-step problems.

Reasoning effort

Use reasoning_effort to trade quality for cost and speed. Models that support it take "low", "medium" or "high"; some also take "minimal".

The samples use the OpenAI SDK set up as in the Quickstart.

response = client.chat.completions.create(
    model="openai/gpt-6-luna",
    messages=[{"role": "user", "content": "How many prime numbers are there below 100?"}],
    reasoning_effort="low",
)
print(response.choices[0].message.content)
print(response.usage.completion_tokens_details.reasoning_tokens)

reasoning_effort is the only reasoning setting HyperFlowSI passes on. A host’s own reasoning options, such as a thinking budget, are removed from the request.

Seeing the reasoning

Some models return their thinking in message.reasoning_content. Others keep it hidden and only report how many tokens it took, in usage.completion_tokens_details.reasoning_tokens.

Billing

Reasoning tokens are charged at the model’s reasoning price where it lists one, and at its output price otherwise, even when you can’t see them. The hold uses the higher of the two. Set max_tokens with room for both the thinking and the answer: if the model runs out while thinking, you pay for the thinking and get no answer.