OpenAI has published a technical breakdown of GPT-Live, its third-generation voice AI system, explaining how the company rebuilt its architecture over six months to remove the lag associated with turn-based conversation.
Unlike earlier systems, which relied on separate "turn detector" models to decide when to respond, GPT-Live uses a full-duplex voice model that can listen and speak simultaneously. When deeper reasoning is required, the system delegates to frontier models such as GPT-5.5 without interrupting the live conversation, OpenAI said.
The company said it rewrote its media frontend and inference logic in Go, replacing a previous Python implementation, which improved frame delivery smoothness so that the new system's 95th percentile latency matched the old system's median. OpenAI also said it developed a protocol called WARP, which cuts media and data startup from six network round trips to one. WARP was designed as a set of open specifications with collaborators from the WebRTC community and is being advanced through the IETF's TSVWG working group.
Before full rollout, OpenAI said it ran a silent production test routing a growing share of ChatGPT Voice sessions through GPT-Live in read-only mode alongside the existing Advanced Voice Mode, which surfaced capacity and geography issues not visible in load testing.
