ChatGPT Advanced Voice Mode rolls out to paid users

Background

For years, talking to a chatbot meant a chain of separate steps: speech was transcribed into text, a text model wrote a reply, and a separate system read that reply aloud. Each hop takes time, and the result feels like a relay race rather than a conversation. GPT-4o, the model behind this rollout, was presented as handling media more directly: it accepts any combination of text, audio, image and video as input and generates combinations of text, audio and image outputs^[4]. OpenAI reported that it could respond to audio inputs in as little as 232 milliseconds, with an average of 320 milliseconds, which it compared to human response time in conversation^[4]. That number is the point. In ordinary talk, people take turns within a fraction of a second; a system that needs several seconds of silence per turn cannot carry a natural back-and-forth.

GPT-4o was launched in May 2024, and voice was part of the demonstration from the start.

What happened

Advanced Voice Mode — real-time spoken conversation using GPT-4o — reached ChatGPT Plus and Team subscribers in September 2024^[1]. It was the consumer-facing form of the audio capability shown earlier in the year, and it was limited to paying subscribers rather than opened to everyone at once.

The voices themselves came from the May 2024 launch, where five were demonstrated: Breeze, Cove, Ember, Juniper and Sky^[1].

Key details

One of those five did not survive the launch period. Sky was disabled on 20 May 2024, after Scarlett Johansson said OpenAI had approached her and she had declined^[1]. The episode shaped how the voice set was discussed publicly, and it is the reason the shipped feature carried a shorter list of voices than the one first shown.

Developers were not far behind consumers. On 1 October 2024, OpenAI introduced the Realtime API, which gave developers similar speech-to-speech abilities^[1]. The pattern is familiar from other model releases: a capability first appears inside a chat product, then becomes something outside developers can build on, and from there spreads into other applications.

Impact

Inside ChatGPT, the change was gradual rather than a single switch. The older Standard Voice Mode was retired on 9 September 2025, when all users moved to ChatGPT Voice^[2]. The naming matters less than the direction: voice stopped being an alternative mode bolted onto a text product and became part of the main interface.

The underlying model kept moving as well. In July 2026, GPT-Live-1 mini replaced Advanced Voice Mode as ChatGPT's default voice model^[3]. Advanced Voice Mode is therefore best understood as a stage rather than an endpoint — the point at which low-latency spoken conversation, demonstrated in a launch keynote, became something ordinary subscribers could use every day, and after which the voice layer continued to be swapped out underneath them.