OpenAI publishes the technical deep-dive of GPT-Live, the system that enables continuous voice in agents. The model listens and speaks in the same stream, without conversation turns, with an architecture built over six months to bring latency below the natural conversation threshold.

Why this matters to you. If you’re evaluating or building voice agents, you now have the documented technical foundation of what’s already in production. On August 2nd, we covered avatarin, which has put 30,000 customers talking to a voice agent based on GPT-Realtime in Yamada Denki stores. GPT-Live describes the architecture underneath that kind of system: what it means to remove turns from voice, what latency you need, what trade-offs it involves.

The difference between turn-based voice and continuous voice is the difference between a walkie-talkie and a phone call. The first forces you to say “ready”, wait, respond. The second lets words overlap, lets you interrupt, lets the other person react while you’re still talking. For an agent assisting a customer on the phone or in a store, this is the threshold that separates a demo from a usable tool.

The document arrives as voice solidifies as an operational channel for agents: Anthropic has extended voice mode to Opus and Sonnet, and OpenAI Presence is betting on voice in the enterprise platform.

In detail

Voice conversation with an AI has a problem that text doesn’t: latency. When you type in chat, a second of waiting is normal. When you talk on the phone, a second of silence after your sentence feels like a void, and two seconds sounds like the other person isn’t listening.

Traditional voice systems solve the problem in phases: they record audio, convert it to text, generate a text response, convert it to voice, play it back. Each phase adds latency, and the turn is rigid: you speak, the AI is silent, then the AI speaks and you’re silent. It’s the walkie-talkie model.

What changes with GPT-Live. The system described by OpenAI eliminates the turn-based structure. The model listens to audio in streaming and generates voice in streaming, in the same continuous flow. This means the agent can start responding before you finish speaking, can be interrupted, and can handle overlaps like a person would on a call. The document describes the architecture built over six months to achieve this, with the explicit goal of keeping latency below the perception threshold.

For the non-engineer. The technical leap is wiring incoming audio directly to outgoing audio, without transcription as a separate step. The model processes voice fragments instead of written word fragments. The architecture must handle the bidirectional stream in real-time, buffer enough not to lose words, and maintain constant latency even when load varies.

The connection with avatarin. The avatarin case in Yamada Denki stores, covered on August 2nd, is the first real-world demonstration of a GPT-Realtime-based voice agent in 24/7 production, with 30,000 customers and 92% stated satisfaction. What was missing was documentation of the underlying architecture. Now it exists: anyone evaluating building a similar system has a technical reference instead of a marketing case study.

The limits of what we know. The document published by OpenAI is a technical deep-dive, but from the summary we know the general approach and results, not the complete implementation details. Specific latency metrics, costs per minute of conversation, and direct comparisons with competing systems (like Claude’s voice mode on Opus and Sonnet) don’t emerge from the summary. What we know is that the system exists, works in production, and OpenAI chose to document its architecture. What we don’t know is how replicable it is by third parties with open tools, and what compute infrastructure it requires to handle the load of mass distribution.

For anyone following the thread of voice as an operational channel for agents, GPT-Live is the technical piece that avatarin took for granted and that Anthropic introduced without documenting. Now the reference exists.