共有:

OpenAI · GPT-Live

You can finally
interrupt the AI.

Voice modes so far waited patiently until you stopped talking — a strict question, then answer, then question. On July 10, OpenAI shipped its full-duplex voice model GPT-Live alongside GPT-5.6, moving the stack toward conversations that allow overlap, back-channels, and interruption.

AI Navigate Editorial2026.07.106 min read

HALF DUPLEX user wait user AI reply AI reply AI reply strict turn-taking FULL DUPLEX user is still speaking AI replies and interjects in parallel overlap and interruption OK
01

The Turn-Taking Wall

What made voice AI
feel so stiff

ChatGPT voice mode, Google's Duplex, Anthropic's voice stack — the awkwardness of AI voice was never really about audio quality or vocabulary. It was about turn-taking. Humans overlap, murmur "mm-hmm," cut each other off, and finish each other's sentences. Ban all of that at the protocol level and the "phone-natural" claim collapses.

Technically, the industry standard was half-duplex: Voice Activity Detection (VAD) marked the end of your utterance, then the model streamed a reply. The AI could not speak while you were speaking, and you could not cut in while it replied. So the moment a real customer said "wait, hold on" mid-response, the dialogue broke.

Before (half-duplex)GPT-Live (full-duplex)
Reply only after user stopsReply is generated while user speaks
Overlap breaks the dialogueOverlap, interruption, back-channel are OK
VAD detects end-of-utteranceStreams both directions in parallel
The "gap" is the giveawaySilences fill in; feels like a live sparring

From an AI that waits for you
to an AI you can interrupt.


02

How It Fits

Two days after the mini
tier, the flagship arrives

The lightweight Realtime API 2.1 mini landed on July 8. The flagship rides the same pipe two days later.

Realtime API 2.1 (foundation · low-latency streaming pipe) 2.1 mini (lightweight, low cost) GPT-Live (full-duplex flagship) added 7/10 added 7/8 one API, one pipe
FIG. Foundation first, then a lightweight tier, then the flagship — all on the same Realtime API.

The point is that 2.1 mini and GPT-Live share the same Realtime API. Builders can route by task: cheap intake bots on mini, dialogues that need real interruption on Live — same pipe, different tier. Two days ago the light tier landed; today the flagship arrives. OpenAI is finishing the voice stack in a single week.

GPT (OpenAI) had frozen its model launches through late June under the new US customer-level approval rule, and only shipped GPT-5.6 on July 9 as the first re-opened case. GPT-Live is the follow-on riding right behind — the "inventory polished during the freeze".

03

By The Numbers

Where the stack
stands

2 days
mini → flagship gap
full-duplex
replies while user speaks
GPT-5.6
the text sibling shipped same day
04

In Practice

Where you feel it

The teams already routing customer calls through AI voice will notice first.

Contact-center calls

A customer's "wait, hold on" no longer breaks the loop — the AI can pick it up, back-channel, and stay in flow instead of resetting.

Voice-driven work

Steering a coding session or draft by voice benefits from being able to cut in mid-answer with "no, the other one." A candidate switch for teams already using text IDEs with voice.

Language / interview practice

Voice AI was always pitched at drilling conversation. Full-duplex adds "how to interrupt" and "how to handle being interrupted" to the drill.


05

Frontier

Who this actually
touches

The reach here is narrow: this matters to people who use or build voice AI. Text chat and batch API users are unaffected today. But for contact centers, voice-UI outsourced spend, and language products, GPT-Live opens a genuinely new option — you can now redesign the turn-taking protocol, not just tune the voice.

The counter-view: full-duplex adds a false-interrupt risk. Well-meant back-channels can cut off users, and overlap can lose intent. This will need tuning in production. The practical read is: run 2.1 mini as your workhorse, and stage in GPT-Live where the dialogue genuinely needs to overlap.