OpenAI has launched GPT-Live, a pair of new voice models that fundamentally redesign how people interact with ChatGPT, replacing the previous Advanced Voice Mode with an architecture capable of listening and speaking simultaneously, much like a human conversation. The models, GPT-Live-1 and GPT-Live-1 mini, are rolling out globally today on iOS, Android, and ChatGPT.com. GPT-Live-1 becomes the default voice model for paid users on the Go, Plus, and Pro tiers, while GPT-Live-1 mini serves free-tier users. OpenAI also plans to offer the models via API, and developers can sign up for notifications.
This release marks the third generation of ChatGPT's voice technology in about two years and represents OpenAI's clearest attempt to turn the chatbot into something that feels less like querying a search engine and more like talking to a colleague. The defining technical advance is the full-duplex architecture, which in telecommunications means both parties can talk and listen at the same time. Applied to AI, it means the model continuously processes incoming audio while generating its own spoken response, eliminating the need to wait for a clean silence gap to determine when the user has finished speaking.
Sponsored Protocol
Full-duplex model eliminates unnatural interruptions and improves dialogue flow
According to OpenAI, GPT-Live continuously processes input while generating output, making interaction decisions many times per second: whether to speak, continue listening, pause, interrupt, or invoke a tool. In practice, the voice assistant can insert conversational acknowledgments like "mhmm," "yeah," or "got it" while the user is still talking, pick up on a natural pause without jumping in prematurely, and handle rapid interruptions without derailing the exchange. The previous Advanced Voice Mode, launched in September 2024, operated on rigid turn-by-turn exchanges where silence detection led to unnatural interruptions, such as in a noisy café or during a thinking pause. GPT-Live is designed to overcome these limitations.
Separation of voice and reasoning layers for greater flexibility
GPT-Live introduces a second structural innovation: it decouples the voice interaction layer from the reasoning layer. When a user asks a simple question, the model handles it directly. For complex requests like web searches or advanced reasoning, GPT-Live delegates the task to a frontier model running in the background—initially GPT-5.5—while continuing to converse with the user asynchronously. This modular design allows upgrading the intelligence of the voice assistant without retraining the voice model itself, with significant implications for enterprise workflows. A voice agent can maintain a natural conversation with a customer while querying databases or performing multi-step reasoning, tasks that introduced several seconds of dead air under the old system.
Sponsored Protocol
From early versions to GPT-Live: three generations of voice AI compared
The first version, ChatGPT Voice in 2023, used a cascaded pipeline: a speech-to-text model (Whisper) transcribed audio, an LLM (GPT-4) generated a text response, and a text-to-speech model converted it back to audio. Each handoff introduced latency and information loss, with a total delay of about 1,700 milliseconds. The Advanced Voice Mode in 2024 collapsed this into a single native audio model but still operated on a turn-based exchange. GPT-Live represents a qualitative leap, eliminating turn-based structure and introducing full-duplex. OpenAI has also remastered the nine distinct voices and implemented safeguards to prevent impersonation of real people, a hot topic after the Scarlett Johansson controversy in 2024 when a voice resembling the actress was pulled after criticism. GPT-Live is designed to move past these controversies, focusing on natural conversation rather than impersonation.
Sponsored Protocol
More than 150 million people use ChatGPT's voice features weekly. New features include visual cards showing weather forecasts, stock data, and maps, three reasoning levels (Instant, Medium, High), and improved background noise handling. Safety tests show significant improvements, with safety scores rising in areas like illicit behavior and self-harm. However, OpenAI acknowledged that risks like emotional reliance require ongoing monitoring.
Sponsored Protocol
Meanwhile, competitors such as Google with Gemini Live, ByteDance with Seeduplex, and Nvidia with PersonaPlex have already launched full-duplex systems. The race to natural voice AI has become table stakes for consumer products, and OpenAI leverages its user base and advanced model integration to maintain a competitive edge. GPT-Live does not yet support video or screen sharing, and API access is not immediate, but the company promises continuous updates.
To learn more about AI and image generation, read the article on Meta using public Instagram photos for AI. Additionally, an in-depth look at full-duplex architectures is available on Wikipedia.
