OpenAI has overhauled ChatGPT's voice infrastructure by introducing GPT-Live, a full-duplex architecture designed to eliminate the latency bottlenecks inherent in traditional turn-based voice-activity detection. Unlike previous "walkie-talkie" models that required discrete pauses for turn detection, GPT-Live utilizes a continuous speech model that can listen and speak simultaneously, allowing for natural interruptions. Architecturally, OpenAI decoupled the audio loop from application logic, creating a dedicated fast path for speech while offloading reasoning, tool calls, and web searches to asynchronous models, such as GPT-5.5 Instant or GPT-5.5 Thinking. To ensure consistent latency, the media frontend and inference logic were rewritten in Go from a Python asyncio implementation, focusing on improving 95th-percentile frame-delivery performance. Furthermore, OpenAI developed the WebRTC Abridged Roundtrip Protocol (WARP) to reduce session startup from six network round trips to a single trip, supplemented by an "Instant Connect" system using UDP packets. This new stack, currently powering ChatGPT Voice for over 150 million weekly users, employs a seamless instance handoff process to manage growing conversation contexts and capacity shifts without interrupting the audio stream.