OpenAI rolls out GPT-Live, a full-duplex voice system for ChatGPT

GPT-Live is OpenAI's answer to the demand for genuinely conversational voice AI. Unlike turn-based voice modes, it is full-duplex — the system can listen while it speaks, letting users interrupt mid-sentence and get an immediate change of direction, which is the core UX gap in prior voice assistants.

The clever mechanism is task delegation: the live voice model handles the real-time conversation while routing anything requiring current information or heavy reasoning to a background frontier model (sources cite GPT-5.5/GPT-5.6). That split keeps latency low for chat while preserving answer quality for hard queries. GPT-Live-1 is the default for paid tiers with a lighter GPT-Live-1 mini for free users, available on web, iOS and Android.

The launch slots into a busy voice week — Alibaba's Qwen-Audio TTS topped the Speech Arena leaderboard, and Axios flagged voice agents as a battleground. It also arrives alongside OpenAI's other consumer moves this week (ChatGPT ads reaching a reported $100M annualized run-rate, WhatsApp availability restored in the EEA). Watch for reliability under the same demand surge Altman warned about, and for how well the background-model handoff hides latency in practice.