OpenAI has launched GPT-Live, a new voice model architecture designed to make interactions with AI feel more natural and conversational. Unlike previous cascaded or turn-based systems, GPT-Live employs a full-duplex architecture, allowing it to listen and speak simultaneously, much like human conversation. This enables quicker back-and-forth, the ability to interject, and naturalistic conversational cues like "mhmm."

Beyond its conversational fluency, GPT-Live delegates complex tasks, such as web searches or deep reasoning, to OpenAI's latest frontier models in the background. This separation allows the core voice model to maintain conversational flow even while the underlying AI performs intensive work. At launch, GPT-Live utilizes GPT-5.5 Instant for immediate responses, with options for medium and high reasoning, and will continuously integrate newer frontier models.

This new architecture addresses limitations of earlier voice AI, which suffered from slow, stilted responses and rigid turn-taking. Human evaluations show GPT-Live-1 and GPT-Live-1 mini are strongly preferred over previous iterations for conversational flow and naturalness. OpenAI plans to roll out GPT-Live to ChatGPT users globally, with upcoming API access for developers and enterprises.