Reasoning Control

For models that support reasoning (chain-of-thought), you can control the depth of reasoning with the reasoning_effort parameter.

Prerequisites

• An API key (get one here)
• OpenAI SDK installed (Quickstart)
• A reasoning-capable model (see Chat & Reasoning models)

What you’ll learn:

• How to tell whether a model supports reasoning control
• How to use the reasoning_effort parameter
• How to disable reasoning for Qwen 3.6 models

Examples of Reasoning Control

These are examples, not a complete list

The table below shows a few common models and how reasoning is controlled on each. It is not an exhaustive list — other models on the platform support reasoning control too, and the catalogue changes over time. For the current list, see Chat & Reasoning models or query the models endpoint.

Model (example)	Reasoning Control
GPT-5, GPT-5 mini, GPT-5 nano	reasoning_effort parameter
Gemini 3	reasoning_effort parameter
Claude 4.6, Claude Opus 5	reasoning_effort parameter
Qwen 3.6	/no_think prompt keyword

Most reasoning models follow one of these two patterns: an OpenAI-compatible reasoning_effort parameter, or a prompt keyword. reasoning_effort is accepted far more widely than the table suggests — GLM, Mistral, Gemma, Nemotron and the T-Cloud-hosted open-source models take it too.

A few models are exceptions worth knowing about:

• Mistral models accept only none and high. Sending low or medium returns HTTP 400 with reasoning_effort=... is not supported by Mistral models.
• Some newer Anthropic models reject reasoning_effort and control thinking differently; the error names the alternative to use.
• Code-specialised models may reject the parameter outright with The requested operation is unsupported.

If a model rejects reasoning_effort, check its entry in Chat & Reasoning models.

Using reasoning_effort

Python

from openai import OpenAI

client = OpenAI()

response = client.chat.completions.create(
    model="claude-sonnet-4.6",
    messages=[
        {"role": "system", "content": "You are a concise and helpful assistant."},
        {"role": "user", "content": "What is the capital of France?"},
    ],
    reasoning_effort="low",  # "low", "medium", or "high"
)

print(response.choices[0].message.content)

The reasoning_effort parameter accepts three values:

• low — Fast responses, minimal reasoning
• medium — Balanced reasoning depth
• high — Deep reasoning for complex problems

Disabling Reasoning for Qwen 3.6

For Qwen 3.6 models, prepend /no_think to your prompt to disable reasoning:

from openai import OpenAI

client = OpenAI()

response = client.chat.completions.create(
    model="Qwen3.6-35B-A3B-FP8",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "/no_think What is your name?"},
    ],
)

print(response.choices[0].message.content)

Next Steps

• Chat Completions — Standard chat API usage
• Function Calling — Connect reasoning models to external tools
• Streaming — Stream reasoning model responses