OpenAI Releases Real-Time Voice Model «GPT-Live»

OpenAI released its new voice model «GPT-Live» simultaneously worldwide on Wednesday, 2025. The model is being progressively made available on iOS, Android, and ChatGPT.com, officially replacing the «Advanced Voice Mode» that ChatGPT has employed to date. The lineup consists of two versions: «GPT-Live-1» and «GPT-Live-1 mini», with GPT-Live-1 becoming the default voice model for paid plans (Go, Plus, Pro), while free users receive GPT-Live-1 mini.

The most significant technical feature of this release lies in what OpenAI calls a «full-duplex architecture». Full-duplex is a telecommunications term referring to «a state in which both parties can speak simultaneously», as in a telephone line. By applying this to AI, the model can now continuously process the user's voice even while generating its own speech output. In other words, there is no need to wait for either party to finish speaking. OpenAI explains on its official blog that «the model makes decisions multiple times per second, determining whether to speak, continue listening, pause, interject, or invoke tools».

The previous Advanced Voice Mode, deployed to paid users in September 2024, operated by determining whether the user had finished speaking based on «the length of silence». This caused problems such as café noise or conversational pauses being misidentified as «end of speech», leading to AI interjecting at unnatural moments. GPT-Live eliminates this dependence on silence detection and is designed to continuously read the flow of conversation from context.

Another structural change is the separation of «voice interaction» and «thinking and reasoning» into distinct layers. This enables updating just the voice model portion or independently enhancing the reasoning model. Enterprise (API) deployment is also planned, and developers can now register for notifications.

This release represents the third generation of ChatGPT's voice capabilities over approximately two years. Previous voice AI has sometimes been described as speaking «like a walkie-talkie, alternating turns», and has been criticized as far from natural conversation. Full-duplex technology differs qualitatively from its predecessors in that it enables inserting «nodding» and affirmations like «yes, yes» that are characteristic of human-to-human conversation while the other person is still speaking.

The naturalness of voice interfaces has long been considered a challenge for AI assistants to become everyday tools. Whether conversational «pauses» and «interruptions» can be handled smoothly significantly affects usability, particularly in professional applications and hands-free environments. Going forward, attention will focus on how developers utilize GPT-Live through the API and how extensively voice-based AI agent applications will expand.