OpenAI rolls out ChatGPT’s advanced voice mode to paid users

OpenAI began rolling out GPT-4o’s Advanced Voice Mode in ChatGPT to Plus and Team subscribers on Tuesday. The feature brings the assistant closer to a phone conversation: it responds with less delay, allows interruptions and can vary its intonation to convey, for example, enthusiasm, calm or empathy.

The launch turns one of GPT-4o’s most striking demonstrations into a real product. OpenAI introduced the model in May. The feature is not yet available to all users or in every country, but it marks a significant shift in how people use a chatbot: voice is no longer an add-on to text, but a primary interface.

A conversation that doesn’t wait for you to finish

Voice interaction already existed in ChatGPT, Siri, Alexa and Google Assistant. However, those systems typically work as a sequence of steps: they first transcribe the audio, then interpret the text and finally generate spoken output. That process introduces pauses and can miss useful signals, such as tone, laughter or the urgency behind a question.

Advanced Voice Mode draws on GPT-4o’s audio capabilities to process and generate speech more directly. In practice, it lets users talk over the assistant to correct it or change the subject without waiting for it to finish responding. It can also adapt the pace and expressiveness of its voice to the conversation.

This is no small matter. With text, users can reread, edit and pause. In a spoken conversation, long silences and rigid turn-taking quickly break the sense of dialogue. Reducing that friction could make ChatGPT more useful while walking, cooking, studying a language or preparing a presentation without having both hands free.

Nine voices and a gradual rollout

OpenAI offers nine preset voices in this version: Arbor, Breeze, Cove, Ember, Juniper, Maple, Sol, Spruce and Vale. Five of them are new. The company restricts the system to voices it has explicitly approved, an important precaution at a time when voice cloning can facilitate impersonation and fraud.

Access is being enabled gradually in the ChatGPT mobile apps for iOS and Android. OpenAI expects to complete the rollout for Plus and Team users this week, although it continues to impose daily usage limits that may vary with demand. Once the available GPT-4o time runs out, the app may switch to a voice experience powered by GPT-4o mini.

The feature is not currently available to paid customers in the European Union, the United Kingdom, Switzerland, Iceland, Norway or Liechtenstein. OpenAI has said it is working to bring it to those markets, but has not set a date.

May’s demo arrives with more controls

GPT-4o’s initial presentation showed an AI capable of holding a fast-paced conversation, detecting apparent emotions and reacting to what it saw through a camera. OpenAI delayed the advanced voice launch, which had been planned for June, to conduct additional safety and reliability testing.

The system raises risks distinct from those of a written chatbot. A response delivered in a persuasive voice can create excessive trust; moreover, an AI capable of accurately imitating a person would have obvious potential for abuse. As a result, the model has been trained to produce only the voices selected by OpenAI and includes detection and blocking mechanisms for attempts to generate problematic content.

The first version also does not include the video and screen-sharing capabilities OpenAI showed in May. The company continues to present them as a planned future expansion.

For users, the real test will be whether the promised naturalness holds up outside a controlled demonstration: in long conversations, amid background noise, with diverse accents and complex questions. For the industry, the move increases pressure on traditional voice assistants, which have spent years promising natural dialogue but still rely largely on short commands and scripted responses.