OpenAI Ships GPT-Live Voice to Desktop, Turning Speech Into Agent Control
ChatGPT Voice lands on macOS and Windows desktops, letting you control your computer and direct multiple AI agents using only your voice.
- ChatGPT Voice is now live on macOS and Windows for Plus, Pro, Business, Edu, and Enterprise plans, rolling out globally today.
- Powered by GPT-Live, a full-duplex voice model that listens and speaks simultaneously, unlike the old turn-based Advanced Voice Mode.
- Voice controls the entire desktop app, including computer use via Codex and multi-agent orchestration in ChatGPT Work.
- Two-layer architecture: a lightweight voice model handles conversation flow while GPT-5.5 handles complex reasoning and search in the background, seamlessly.
- Part of a broader consolidation: Codex, Chat, and Work are now unified in one ChatGPT desktop app, with voice as the top-level control layer.
- Security surface is larger: voice-triggered computer use on real machines warrants careful review before broad enterprise deployment.
OpenAI has shipped ChatGPT Voice to the desktop app, rolling it out globally on macOS and Windows for Plus, Pro, Business, Edu, and Enterprise plans. The headline feature is voice-controlled computer use and multi-agent orchestration, all driven by a single spoken command.
The engine underneath: GPT-Live
The feature runs on GPT-Live, the voice model architecture OpenAI launched on July 8, and it represents a genuine architectural departure from what came before.
Previous voice systems were half-duplex: the model waited for you to stop talking, then responded. Advanced Voice Mode processed and generated audio within a single model but still operated through discrete turns, meaning even a brief pause or background noise could be mistaken for the end of a turn.
GPT-Live runs on a full-duplex architecture that processes input and output simultaneously. It is also a two-layer system: a continuous interaction layer manages the conversation, while a delegation layer quietly hands complex tasks to GPT-5.5 in the background. The system makes decisions multiple times per second about whether to speak, listen, pause, interrupt, or invoke tools, and incorporates natural cues like "mhmm" and "got it" to signal engagement. The result feels less like querying a model and more like talking to a colleague who is already pulling up the file while you finish your sentence.