OpenAI launches GPT-Live-1 voice models for live conversation

OpenAI released new conversational models, called GPT-Live-1 and GPT-Live-1 mini, today. The company says the models are full-duplex, so they can speak and listen at the same time, which lets users interrupt naturally and enables features such as live translation and longer hands-free conversations.

What did OpenAI release?

OpenAI released two new voice models: GPT-Live-1 and a smaller GPT-Live-1 mini, and it will make GPT-Live-1 mini the default replacement for ChatGPT's current Advanced Voice Mode while reserving the larger GPT-Live-1 for paid tiers. The models are described as full-duplex conversational systems that replace the prior pipeline of separate speech-to-text, a large language model, and text-to-speech.

The company said the new models will forward queries to its latest text models, including GPT-5.5, for tasks such as search, reasoning, or agentic capabilities while maintaining the live conversation. Product lead Atty Eleti said he has had 30- to 40-minute-long conversations with the voice feature during walks, and the company noted that more than 150 million people talk to ChatGPT using features like Voice and Dictation.

How do the GPT-Live models work differently?

GPT-Live models operate as continuous, bidirectional voice agents that can listen while speaking, absorb ongoing context and pause until addressed, instead of strictly processing discrete turns. OpenAI said this full-duplex design prevents interruptions that used to occur when the assistant would speak over users and aims to provide better turn-taking and longer conversations.

The new voice mode can also present some information visually because it has access to newer GPT text models, and OpenAI demonstrated live translation in the demo. The company acknowledged limitations: during the demo the Hindi translation had a heavy American accent and sounded slightly bookish, and it did not specify which languages the mode is optimized for. OpenAI also emphasized built-in safeguards to give age-appropriate responses to teens and to provide resources if a conversation turns to topics like self-harm.

Why it matters

Full-duplex voice removes a core friction in spoken interfaces: forced turn-taking. By letting the assistant listen while speaking and route complex requests to text models like GPT-5.5, OpenAI is positioning voice as a more capable interface for longer, goal-directed interactions rather than short queries. That could shift how people use assistants for tasks that require extended context, chaining, or multiturn problem solving.

The change also tightens the competition: Apple and Amazon have updated assistants to be more conversational, and startups are pushing expressive, task-focused agents. OpenAI highlighted the scale of voice use inside ChatGPT, saying over 150 million people use Voice and Dictation, which matters because adoption can drive developer and product focus toward hands-free workflows.

What to watch

Watch whether the live translation and non-English voice quality improve beyond the demo, and whether OpenAI specifies which languages GPT-Live-1 is optimized for. Also watch rollout signals: GPT-Live-1 mini replaces Advanced Voice Mode by default, while the larger GPT-Live-1 is promised for paid tiers, and reports suggest the company may explore AI earbuds later this year though no hardware details were provided.