ChatGPT Can Now Hear You and Do Things on Your Computer

OpenAI has added voice control to its ChatGPT desktop app. You can now speak to ChatGPT and ask it to carry out several steps in a row, and it will talk back to you when it needs more information. The feature became available on July 23, 2026, with the official announcement following on July 24 (TechCrunch).
This is powered by OpenAI's newer voice technology, called ChatGPT-Live, which launched earlier in July 2026. Voice in the desktop app works with ChatGPT Work and Codex, which is OpenAI's tool for writing and managing code. The app can also look up websites and open applications on its own. On Mac computers, a feature called Appshots lets ChatGPT Voice see what is on your screen, including text descriptions hidden inside images (TechCrunch).
The feature does more than just let you chat by voice. OpenAI shared a demo video showing a developer speaking a single command that created a new project thread, submitted a proposed code change, and tracked down the cause of a software bug. The desktop version speaks out loud when it needs you to answer a question, so you can keep working without going back to the keyboard each time (TechCrunch).
For Codex users, Voice also works from the iPhone app. That means a developer can start a voice session with Codex from their phone, even if they are not at the computer where the work is actually running (TechCrunch).
This is not the first time OpenAI has offered voice in the Mac desktop app. An earlier version was taken away on January 15, 2026 (OpenAI Help). The phone version of ChatGPT Voice that launched with the ChatGPT-Live family felt smoother and handled interruptions better, but it could not actually do things on your device. The desktop version fixes that by combining voice with tools that can control the computer.
The release comes as a rival company, Anthropic, is also improving voice features. Around July 23, 2026, Anthropic updated its assistant Claude so voice commands can complete tasks in apps like Gmail, Calendar, Slack, Notion, and Canva (TechCrunch). Both companies are heading toward the same goal: letting you speak a task and have the assistant carry it out across whatever apps are needed, rather than just talking with you.
The difference between the phone and desktop versions is important. The phone version focused on sounding natural and handling interruptions smoothly. The desktop version focuses on getting tasks done, running several steps from one spoken command and using what is on your screen to make better decisions. The Appshots feature on Mac is especially useful: by reading what is on your screen, the assistant can base its answers on what you are actually looking at, so you do not have to describe everything out loud.
A demo where everything works in one smooth command is impressive, but there is a known pattern where these kinds of automated, multi-step tasks can fail in unexpected ways when conditions are not exactly as expected. People who want to use this for real, time-sensitive work should learn how it fails before depending on it.
The iPhone access for Codex could change how developers work. Instead of checking in on a coding task every so often, they could keep a voice conversation going with it from anywhere, turning occasional check-ins into an ongoing dialogue.
The competition with Anthropic adds pressure on both quality and reach. Claude's voice mode currently works with everyday productivity apps like Gmail and Slack. OpenAI's desktop approach targets developers through Codex and by reading the screen. The two strategies do not fully overlap yet, but both are moving toward the same expectation: you speak a task, and the assistant does it across whatever tools are needed.
The earlier desktop Voice feature was gone for about six months between its January retirement and this release. That gap suggests OpenAI decided the previous version was not good enough and chose to start over with the newer ChatGPT-Live technology rather than make small improvements. The result pairs a better speech model with the computer-control tools the desktop app already had, making it a different offering from what came before.


