OpenAI has officially launched its new generation of voice models, GPT-Live-1 and GPT-Live-1 mini, representing a comprehensive upgrade to ChatGPT's voice capabilities.
The new models employ a "full-duplex" architecture, enabling them to listen and speak simultaneously. This allows users to interrupt the AI's responses at any time, creating an interaction style that more closely resembles a real human conversation.
At the launch event, OpenAI's research lead, Kundan Kumar, explained that the previous advanced voice mode in ChatGPT used a turn-based dialogue system. The model had to wait for the user to finish speaking before responding, and brief pauses were often misinterpreted as the end of speech, leading to unnatural interruptions. GPT-Live solves this issue by continuously processing input while simultaneously generating output, making decisions multiple times per second on whether to speak, continue listening, or interrupt. During conversations, the model will use brief acknowledgments like "hmm," "right," or "got it" to signal it is actively listening. Users can also ask it to slow down its speech or remain quiet while awaiting further instructions.
Another key improvement in the new models is the separation of the voice interaction layer from the reasoning layer. When a question requires a web search or more complex reasoning, GPT-Live will delegate the task to the backend GPT-5.5 model while maintaining a smooth conversational flow. Paid subscribers can choose between three reasoning levels—Instant, Medium, and High—to balance speed and intelligence based on their needs.
Furthermore, GPT-Live supports real-time translation and can display information such as weather, stock prices, and sports scores through visual cards. OpenAI stated that GPT-Live-1 will become the default voice model for Go, Plus, and Pro subscribers, while free users will have access to the GPT-Live-1 mini variant. Both models are being rolled out globally starting today to iOS, Android, and web users, with API access to follow at a later date. With over 150 million people using ChatGPT's voice and dictation features weekly, this upgrade is expected to significantly expand the use cases for voice-based interaction.