Talking to a live person — ALMOST ALL of it (sorry)

The new model runs on a full-duplex architecture, meaning it can listen and speak at the same time instead of processing messages one after another. It doesn’t jump in at every pause, can back-channel with cues like “mhmm” or “got it,” go quiet when you need a moment to think, and pull in web search while reasoning in the background through GPT-5.5 for harder questions — all without breaking the flow of conversation.

In short, the neural network has officially learned not just to answer, but to actually hold a conversation the way a person would, including handling interruptions, natural pacing, and live translation.

Full-duplex is the architectural detail doing most of the work here. Previous voice modes operated in a strict turn-taking loop — listen, process, respond — which is exactly why they felt mechanical even when the underlying text model was capable of natural-sounding replies.

GPT-Live is already rolling out globally in ChatGPT across iOS, Android, and the web. GPT-Live-1 will become the default voice model for paid users on Go, Plus, and Pro plans, while GPT-Live-1 mini will serve as the default for Free users.

In internal human evaluations of multi-minute conversations, both versions were strongly preferred over the previous Advanced Voice Mode, and OpenAI says it plans to bring GPT-Live to the API for developers soon.

For affiliates and content teams building voice-driven products, customer support flows, or interactive demos, the gap between GPT-Live and the previous Advanced Voice Mode is worth testing directly once API access opens.