OpenAI launched GPT-Live, a new family of voice models capable of listening and speaking at the same time, on Wednesday.
The company said it is rolling out two versions — GPT-Live-1 and GPT-Live-1 mini — to users globally. GPT-Live-1 mini will replace the current Advanced Voice Mode in ChatGPT by default, while users on paid tiers will have access to the larger GPT-Live-1 model, the company said.
Unlike traditional voice systems, the new models operate in full-duplex mode — handling speech input and output at the same time — which opens the door to natural interruptions and capabilities such as live translation, TechCrunch reported. Audio processing in the old system worked as a relay: spoken words were first transcribed, then fed to a language model, then converted back to speech before reaching the user.
At a press briefing, OpenAI described shortcomings in the older system that the new models are meant to fix, pointing to cases where the assistant cut off users and fell short on question-answering. When a query calls for search, reasoning, or agentic capabilities, the system hands it off to a text model such as GPT-5.5 in the background without pausing the voice interaction, the company said.
The new voice mode is also designed for longer conversations. ChatGPT Voice product lead Atty Eleti said during a press briefing that he had held 30- to 40-minute conversations with the feature during walks.
Demonstrations also highlighted the model's ability to hold back and quietly take in what is being said until it is needed. Because the voice mode draws on newer GPT models, it can surface certain responses visually rather than through speech alone.
The company said more than 150 million people use ChatGPT through voice features such as Voice and Dictation. OpenAI said it sees voice as a potential primary interface for complex computing work.
"Over time, we think this will also unlock the ability to use voice as a kind of primary interface to computing, and to manage increasingly complex long-running agentic work," Eleti said.
OpenAI said the models include safety features, such as guardrails that change responses for younger users and suggest support resources if sensitive topics like self-harm are mentioned. The company also said it is not presenting the new voice mode as an AI companion.
Earlier this year in May, the company released a trio of audio models targeted at developers building voice-driven applications, with a focus on making those agents sound more natural in conversation, Reuters reported.