GPT-Live: Full-Duplex Voice Model for Natural Conversations

Previous voice AI systems forced an unnatural tradeoff: either slow, turn-based responses from cascaded models, or slightly smoother but still rigid interaction where the model waited for silence before replying. The original ChatGPT Voice chained three models—speech-to-text, LLM, text-to-speech—losing information and adding latency. Advanced Voice Mode improved by processing audio in a single model, but its turn-based design meant the model could interrupt at unnatural moments due to silence detection. Users got polite but stilted conversations, not the fluid back-and-forth of a real dialogue.

GPT‑Live tackles both problems with two architectural changes. First, it uses a full-duplex architecture for continuous interaction: the model processes input while generating output, making decisions many times per second about whether to speak, listen, pause, or interrupt. This enables natural acknowledgments like “mhmm,” quick interruptions, or comfortable silence. Second, it decouples the continuous interaction layer from deeper reasoning by delegating complex tasks—web search, scientific reasoning, agentic work—to a frontier model like GPT‑5.5 in the background while GPT‑Live keeps the conversation flowing. At launch, GPT‑Live‑1 and GPT‑Live‑1 mini use GPT‑5.5 Instant; a medium and high variant use the thinking model with adjustable reasoning effort.

The result is a voice experience that human evaluations strongly prefer over Advanced Voice Mode across measures of turn-taking, interruptions, and naturalness. GPT‑Live‑1 also substantially outperforms Advanced Voice Mode on GPQA (expert-level science reasoning) and BrowseComp (agentic web search). OpenAI is rolling out GPT‑Live globally to ChatGPT users today, with API access coming soon. For builders, the key insight is that separating real-time interaction from deep reasoning allows both to excel—and the safety work in the system card shows they’ve thought about emotional reliance and real-time safeguards. This architecture points toward a future where voice is not just a frontend but a persistent, collaborative interface.