GPT-Live Separates Speech From Deeper Reasoning

GPT-Live streams voice continuously, delegates reasoning in the background, and lays the groundwork for a real-time interaction API.

GPT-Live is OpenAI’s third-generation voice system. Its model operates in full duplex, allowing it to listen and speak at the same time without waiting for a separate detector to determine whether the user has finished speaking.

Audio streams continuously between the device and the model through a dedicated path. Deeper reasoning, tool calls, and application-specific policies are handled separately. This allows GPT-Live to consult a model such as GPT-5.5 in the background while keeping the conversation moving.

OpenAI replaced an earlier Python implementation with a system written in Go. According to the company, the new architecture’s 95th-percentile frame-delivery performance now matches the median performance of the previous system. Transport relies on WebRTC, with mechanisms designed to handle packet loss and connection fluctuations.

The WARP protocol reduces the startup of a WebRTC session from six network round trips to one. Another mechanism, Instant Connect, prepares certain parameters before the connection begins, allowing the server to respond as soon as it receives the first audio packet.

For longer conversations, the system prepares a new model instance in parallel, transfers the current context to it, and switches over once it is ready. The same method allows the conversation history to be compacted without interrupting the voice stream or rebuilding the state during the exchange.

This architecture already powers ChatGPT Voice, including its computer-control and agent-coordination features in the desktop app. It is also expected to serve as the foundation for a future GPT-Live API.