Reasoning Support in LLM Responses: ChatGPT, Claude, and Gemini

We’re excited to announce that the LLM Responses API now supports reasoning for specific models. Now, you can explore the step-by-step thought process of popular LLMs and gain deeper insights into their behavior.

What’s new

Here are key changes we made in LLM Responses to support reasoning:

• use_reasoning parameter was added to the Claude and Gemini LLM Responses endpoints. It enables reasoning mode in these LLMs. Note that each model has specific requirements and limitations when using this parameter. See the documentation for details.
• The response’s items array now includes the reasoning item. The new item contains the model’s step-by-step reasoning chain generated before the final answer. Supported for ChatGPT, Claude, and Gemini reasoning models.
• reasoning_tokens field was included in the result object. This field indicates the total number of tokens consumed to generate the reasoning content.

How it works

For Claude and Gemini, set use_reasoning to true in your request to enable reasoning mode. The model will perform reasoning before generating a response, and the reasoning item will appear in the items array when reasoning data is available. For ChatGPT, reasoning is applied automatically without any additional parameters.

Note that use_reasoning is supported only for reasoning-capable models – refer to the Models endpoints for a list of compatible models.

Why it matters

Access to the reasoning chain allows you to go beyond the final output and understand the logic behind it. In practice, this enables you to:

• Analyze model decision-making. See how the model interprets a query, what assumptions it makes, and how it structures its response.
• Understand how LLMs evaluate content relevance. Explore which factors and sources the model considers when deciding what to include in a response.
• Compare model behavior across LLMs. Evaluate how different reasoning models approach the same query to understand specifics of data search and response generation.

Check out the updated documentation for ChatGPT, Claude, and Gemini LLM Responses endpoints and start exploring reasoning data today.