OpenAI launches GPT-Live full-duplex voice in ChatGPT

Background

Spoken conversation is harder to automate than it looks. Most voice assistants work in turns: the system listens, decides the user has finished speaking, composes an answer, speaks it, and only then starts listening again. Engineers call this half-duplex. The result is the familiar walkie-talkie rhythm — the user has to wait for the model to stop, the model cannot be interrupted cleanly, and a pause or a bit of background noise can be mistaken for the end of a turn. ChatGPT's earlier voice experience, Advanced Voice Mode, belonged to this turn-taking tradition and was the product's default voice model ^[1][2].

What launched

On 8 July 2026 OpenAI launched GPT-Live, a new generation of voice models for ChatGPT ^[1]. The defining change is that the models are full-duplex: they can listen and speak at the same time instead of alternating between the two ^[1]. Two models arrived together. GPT-Live-1 mini replaced Advanced Voice Mode as ChatGPT's default voice model, while GPT-Live-1 was made available to paid users ^[1][2]. OpenAI described GPT-Live as its smartest voice model yet, said it had begun rolling both models out to ChatGPT users globally, and noted that API access would follow ^[3]. The launch came a day before GPT-5.6 and ChatGPT Work ^[2].

How it works

Full duplex means the audio path runs in both directions at once, much as two people on a phone call can both talk and hear without taking formal turns. Because the model keeps listening while it speaks, a user can start talking mid-answer and be heard, and the system has acoustic information about the conversation continuously rather than only in the gaps between turns.

The voice model is also not expected to do all the work itself. For web search, deeper reasoning or complex work, it delegates to a frontier model behind the scenes and keeps the conversation flowing while that work is under way ^[3]. The design separates the fast, conversational layer from the slower one that looks things up or reasons through a hard problem, so that the user is not left sitting in silence while the heavier model works.

Watermarking

On 31 July 2026, supported audio generated with GPT-Live began carrying SynthID watermarks ^[3]. A watermark of this kind is a signal embedded in the generated audio itself, intended to travel with the file as a marker that the audio was produced by AI rather than recorded from a person.

Why it matters

Voice is the interface where latency and turn-taking are most obvious to users, and full duplex is a direct attempt to fix that. The two-model structure matters too. Because GPT-Live-1 mini took the default slot, it is the model most ChatGPT users will hear, while the larger GPT-Live-1 is reserved for paid subscribers ^[1][2]. Defaults tend to decide what a product feels like to the majority of people who never change a setting.

The promise of API access extends the same voice behaviour beyond ChatGPT itself, to developers building their own conversational applications ^[3]. The timing also says something about how the release was positioned: it landed immediately before GPT-5.6 and ChatGPT Work, part of a closely spaced sequence of announcements rather than a standalone event ^[2]. And the addition of SynthID watermarking to supported generated audio, three weeks after launch, connected the voice release to the broader question of how synthetic speech is labelled once it becomes hard to tell from a recording ^[3].