OpenAI Launches GPT-Live in 6 Months, Enabling Full-Duplex Voice AI

3 articles · Updated · OpenAI · Aug 3

OpenAI said GPT-Live is its third-generation voice system and replaces turn-based handoffs with a full-duplex model that can listen and speak simultaneously, aiming for more natural, sub-second conversations.
The redesign removes the turn detector from the audio path and splits live media from tool use and deeper reasoning, letting GPT-Live call frontier models such as GPT-5.5 asynchronously without stalling speech.
Six months of engineering work also shifted media and inference services to Go from Python asyncio, with OpenAI saying the new system’s p95 frame-delivery smoothness matches the previous system’s p50.
WebRTC-based transport, plus OpenAI’s WARP and Instant Connect protocols, cuts startup delay enough that a session can begin with a single UDP packet before standard signaling fully completes.
The architecture already powers expanded ChatGPT Voice features, including computer control and agent coordination in the desktop app, and is set to underpin a forthcoming GPT-Live API.