VOL 01 · EDITION 01 8 guides 1 new model 0 ads (yet)

The complete guide to GPT-Live, OpenAI's new voice model

OpenAI just shipped its first true full-duplex voice model. GPT-Live Hub walks you through what changed, how to enable it, what it costs, and what to expect from the API.

What's new today

Latest update — 2026-07-10. We've added a free-tier deep-dive, expanded the language support guide, and verified pricing against OpenAI's official page. If you just want to know whether to switch, read the main guide. If you want the day-one checklist, jump to how to enable.

Start here

Why GPT-Live matters

GPT-Live is OpenAI's first real-time voice model that can listen and speak at the same time. It backchannels naturally, pauses when you pause, and offloads hard questions to GPT-5.5 mid-conversation. If voice is part of how you use ChatGPT, this is the biggest upgrade since ChatGPT itself.

Three shifts are worth pointing out. First, the interaction model: full-duplex replaces turn-based, which removes the awkward "I haven't finished yet" pause that broke older voice modes. Second, voice is decoupled from reasoning. The voice layer can hand off to GPT-5.5 or Web Search mid-conversation without pausing, so research-heavy prompts are no longer hostile to voice. Third, the safety layer is graduated — gentle redirection, on-screen warnings, or a hard cutoff — instead of a single refusal style. Each of these individually would be a small upgrade; together, they change what voice mode feels like day to day.

For the deep dive, start with the main guide. For the daily user, jump straight to how to enable it. For developers, the API page covers what we know so far.

Who this site is for

Daily ChatGPT voice users

You already talk to ChatGPT and want to know whether to switch, when, and what trade-offs come with each plan. Start at the overview.

Buyers weighing Plus vs Free

You're unsure if ChatGPT Plus is worth $20 for the new voice model. Read the pricing page and the free-tier explainer first.

Developers planning a voice app

You want to ship a real-time voice agent. The API page covers release timing, endpoint shape, and how to join the waitlist.

Bilingual and multilingual users

You speak two or more languages and want to know which Tier 1 you should rely on. The language support page has the accent-quality tier list.

What changed under the hood

Compared with the old Advanced Voice Mode, the biggest architectural shift is decoupling. The voice layer used to be tightly coupled to the chat layer; that's why a hard question used to break voice mode. GPT-Live separates them: the voice layer handles streaming audio in and out, while the reasoning layer (GPT-5.5 by default) handles hard questions and tool use. A small scheduler decides which one should be active at any moment, and the user hears the result as one fluid voice session.

The other quiet improvement is the prosody model. GPT-Live introduces natural backchannels ("mm-hm," "right," "got it") and dynamic pacing that adjusts to context rather than stretching the audio. The result sounds less like a customer-service bot and more like a person who happens to read every book very fast.