OpenAI has released GPT-Live, a new family of voice models built on full-duplex architecture, meaning they can listen and speak at the same time, enabling natural conversational flow with backchannels, interruptions, and real-time turn-taking.
The models, available in ChatGPT starting today, represent a significant architectural departure from both the original cascaded ChatGPT voice system and the more recent Advanced Voice Mode. GPT-Live uses a single model that processes input continuously while generating output, making interaction decisions many times per second: whether to speak, listen, pause, interrupt, or invoke a tool.
Two architectural innovations
The first is full-duplex processing. Previous voice systems were turn-based: one party speaks, then the other responds, with silences marking the boundaries. GPT-Live eliminates this constraint. The model can add brief backchannels (“mhmm,” “yeah”) while the user is still talking, can be interrupted mid-sentence, and can pause and resume naturally. It also supports live translation between languages.
The second is delegation. When a user asks something that requires web search, deep reasoning, or agentic capability, GPT-Live hands the task off to GPT-5.5 running in the background while maintaining the conversation flow. The user hears a quick acknowledgment (“one sec, still with you”) and the answer arrives without the voice interaction breaking.
Evaluations
In human preference tests with 5 to 10 minute conversations, GPT-Live-1 and GPT-Live-1 mini were strongly preferred over Advanced Voice Mode across all measured dimensions: overall preference, turn-taking quality, interrupt handling, conversational flow, and naturalness.
On automated benchmarks, GPT-Live-1 substantially outperformed Advanced Voice Mode on GPQA (expert-level science reasoning), BrowseComp (agentic web search), and the tau3-Voice Telecom benchmark for multi-turn telecom support tasks.
GPT-Live-1 (instant) and GPT-Live-1 mini use GPT-5.5 Instant as their background reasoning model. GPT-Live-1 Medium and High use GPT-5.5 Thinking with configurable reasoning effort. Video, screen sharing, and full multilingual parity are not available at launch but are expected in future updates. An API release is planned.
Sources: MarkTechPost (July 8)