Uncovering Hidden Features of the OpenAI Computer Use Preview Model for Efficient Usage

Published On Mon Mar 31 2025

Incorrect API docs for computer use preview - Documentation ...

I have been trying to run the computer-use-preview model via the OpenAI SDK. I have read the documentation.

At this moment, the documentation states that for my task, I wanted to start with an input image of a screenshot as my initial state. So, naturally, I uncommented the code, tried to run it, but I got an error about an invalid type. This left me confused, and I went straight to look at the repository with examples: simple_cua_loop.py.

However, I could not figure out how to start the computer-use-preview model with both a text prompt and my own screen. After some trial and error, I finally found the correct request that looks like this:

Turning first-time academic misconduct into a teachable moment

I hope someone will find this comment useful because when I needed help, I found zero posts about using computer-use-preview from OpenAI. In the content list, there is also a type called "text" that can be used so that it appears that the single user is sending both text and input_image. And with multiple images also.

Thanks for the correction! I spent half a day trying to understand how to at least send a request without errors, so I never thought about optimizing it.

OpenAI also launched an open-source Agents SDK, available ...