OpenAI announced on July 23 (U.S. time) that it has updated the ChatGPT desktop app to add ChatGPT Voice, a way to operate the app by speaking. Users can give spoken instructions to AI agents and get work done on their computer. It runs on GPT-Live, a new family of voice models released earlier this month, and is available first on the macOS and Windows apps.
A Desktop Version That Can Actually Do the Work
Until now, ChatGPT's voice features centered on conversation—asking questions and getting answers. The new desktop version lets you issue complex, multi-step instructions by voice and answer back when ChatGPT needs your input. It has grown from a simple voice chat into a way to move real tasks forward by speaking.
ChatGPT Voice works with both ChatGPT Work, aimed at everyday business tasks, and Codex, the coding assistant. It can also call on "computer use" skills to look things up across websites and apps on its own, so you can hand off everything from research to execution by voice alone. In a demo video, OpenAI showed a developer issuing a single spoken command—create a new thread, open a pull request, and find the root cause of a bug—and ChatGPT carrying out the whole sequence.
Built on GPT-Live, a Model That Listens and Speaks at Once
The feature is powered by GPT-Live (ChatGPT-Live), the voice model family OpenAI released earlier this month. Conventional voice assistants take turns: they wait for you to finish speaking before they respond. GPT-Live rethinks that with a full-duplex approach, listening and speaking at the same time. It can slip in brief acknowledgments like "got it" without cutting you off, while heavier, slower reasoning is handed to background models. The result balances natural conversation with the ability to handle demanding work.
On macOS, It Can See Your Screen—Aimed at Developers
The macOS version adds a screen-awareness feature called Appshots. With your permission, ChatGPT Voice can reference the frontmost window you have open and what it shows, including alt text. Because it can take your local files and code structure into account, the intended use is a developer talking through a problem out loud while agents carry out the actual work in parallel. It feels close to discussing something aloud while looking at your screen, and you can keep making progress with your hands off the keyboard.
For Codex, OpenAI says you can also operate it by voice while away from your desk, through remote access from the iOS app.
How It Differs from the Phone Version, and Anthropic's Move
The smartphone version of ChatGPT Voice improved the smoothness of conversation and how it handles being interrupted, but it was not built to actually take actions on the phone. The new desktop version goes a step further by committing to executing tasks by voice—the key difference.
OpenAI is not alone in pushing assistants that get work done by voice. Anthropic has also revamped the voice mode of its assistant Claude, letting users choose among the Opus, Sonnet, and Haiku models and handle tasks by voice inside connected apps such as Gmail, Google Calendar, and Slack. A clear new axis of competition is taking shape: assistants you can direct by voice to move real work forward.
Summary
OpenAI added ChatGPT Voice to the ChatGPT desktop app, letting users direct ChatGPT Work and Codex agents by voice. It is built on GPT-Live, which listens and speaks simultaneously, and on macOS it can reference what is on your screen. The ability to give multi-step spoken instructions—and even hand off development work—is what sets it apart from before. The competition over assistants that carry out real tasks by voice, including Anthropic's efforts, is heating up.
