OpenAI released ChatGPT Voice for its Mac and Windows desktop apps on July 23, 2026, extending the full-duplex GPT-Live model from phones to the machines where people actually do their work. The feature lets users control their computer and direct multiple agents running in ChatGPT Work or Codex without touching the keyboard, and it began rolling out globally the same day to Plus, Pro, Business, Edu, and Enterprise subscribers.
"ChatGPT Voice is now in the desktop app," OpenAI wrote in its announcement. "Control your computer and direct multiple agents running in ChatGPT Work or Codex, using just your voice. It's powered by GPT-Live, so it can speak, listen, and coordinate work in the app at the same time."
Key Highlights
- ChatGPT Voice landed on macOS and Windows desktop apps on July 23, 2026, rolling out globally the same day.
- It runs on GPT-Live, the full-duplex voice model OpenAI shipped to mobile and the web on July 8, which listens and speaks simultaneously instead of waiting for a turn to end.
- Voice works across all three surfaces of the desktop app — Chat, ChatGPT Work, and Codex — so a spoken instruction can spawn, monitor, and redirect background coding agents.
- On macOS, an optional Screen context permission lets ChatGPT take appshots of the frontmost window and use them as conversation context.
- Available to Plus, Pro, Business, Edu, and Enterprise plans, with Enterprise and Edu accounts receiving a two-week early-access window.
Details
The technical foundation is GPT-Live, the model OpenAI introduced earlier in July. Unlike traditional voice assistants that transcribe, think, then speak in sequence, GPT-Live operates in full duplex: it can talk while listening, handle interruptions mid-sentence, and hand off harder reasoning or web searches to a separate frontier model running in the background. That last property is what makes the desktop version more than a dictation shortcut — the voice thread stays responsive while heavier work executes elsewhere.
Inside the desktop app, that translates into task delegation. A spoken request can open a separate thread for a task, then the same conversation can check on that thread's progress and send follow-up instructions without the user ever switching windows. Users can bind a configurable hotkey under Settings, then Voice, to summon ChatGPT Voice while working in another application, or click the Voice button in the app itself.
Mac users get one extra capability. With Screen context enabled, ChatGPT captures appshots of whichever window is in focus and folds them into the conversation, so asking about "this error" or "this design" resolves against what is actually on screen. Windows does not yet have the equivalent.
OpenAI documented practical limits alongside the launch. Voice usage is metered in rolling five-hour windows and draws from an allowance separate from standard Codex usage. Only one voice chat can be active across the app at a time. Permissions are inherited from whichever environment the conversation is operating in — a voice session in Codex has exactly the access that Codex thread already had, not more.
Impact
For developers, this is the missing input layer on a stack OpenAI has been assembling piece by piece. The company merged Codex into ChatGPT on July 9 to create ChatGPT Work; Codex counted roughly 5 million weekly users at the time of the merger and has since reported around 10 million. Voice control turns a fleet of background coding agents into something a person can supervise conversationally rather than by cycling through tabs.
Early users on X described the shift in exactly those terms. Developer Jonathan Roomer noted that the release only works because several prior components matured together: "Computer use had to become much faster. Codex had to be able to create, monitor and intervene in other threads. Voice had to become natural enough to have a proper conversation." Another user reported describing a bug out loud and watching the system locate the GitHub issue, reproduce the fault, ship a fix with regression coverage, and open a pull request.
The competitive framing is direct. Codex is OpenAI's answer to Anthropic's Claude Code, and voice-driven agent orchestration is a differentiator neither rival had shipped on the desktop. It also lands eight days after xAI announced its own full-duplex voice work, with a notable asymmetry: xAI is selling a no-code Voice Agent Builder to developers at roughly five cents per minute of audio, while GPT-Live remains locked inside ChatGPT with no public API.
What's Next
The obvious open question is whether OpenAI exposes GPT-Live through its API. Until it does, the model's quality is a demonstration rather than a building block, and teams wanting full-duplex voice in their own products will keep looking elsewhere. Also unresolved: whether Windows gains a screen-context equivalent to appshots, and how the five-hour usage windows hold up for developers who leave a voice session open across a working day.
Source: OpenAI