OpenAI drops ChatGPT Voice on desktop — and it can now take action

There’s a specific kind of frustration that comes from talking to a voice assistant that can understand you but can’t actually do anything. OpenAI just made that particular pain point obsolete on the desktop. On Thursday, the company announced it has baked its ChatGPT Voice feature directly into the macOS desktop app, and this isn’t just a port of the smartphone version. This one can reach into your machine and work.

The update taps the newly launched ChatGPT-Live voice models, but the real muscle comes from its integration with ChatGPT Work and Codex. The assistant can now use computer-use skills to browse the web, dig through apps, and even leverage macOS Appshots to see what’s on your screen, including reading alt-text. In a demo, a developer rattled off a multi-step command—create a thread, make a pull request, find the root cause of a bug—and ChatGPT Voice executed it as a single workflow. That’s a leap past smoother conversations and interruption handling; it’s agency.

This move puts OpenAI in a more direct feature race with Anthropic, which recently updated its own Claude voice mode. Claude already connects to the Opus, Sonnet, and Haiku models to complete tasks inside Gmail, Calendar, Slack, Notion, and Canva. But OpenAI’s approach feels more system-level, at least on macOS for now, with that ability to see and control the screen directly. Users can even access the voice features in Codex remotely through an iOS app, which hints at a cross-device strategy that isn’t fully baked yet but is clearly on the roadmap.

I’m watching to see how well this actually handles the messy reality of desktop interfaces. Voice commands that string together multiple actions sound great in a demo, but the tolerance for error when an AI is clicking around your development environment is basically zero. The fact that ChatGPT now asks for input mid-task is a small but crucial design choice—it suggests OpenAI knows it can’t just let this thing run blind.