OpenAI's GPT-Live-1 mini brings natural full-duplex voice to ChatGPT

OpenAI today launched GPT-Live-1 and GPT-Live-1 mini, conversational AI models that process speech and listen at the same time—a full-duplex capability. Unlike the previous pipeline that combined separate speech-to-text, LLM, and text-to-speech models, these new models handle turn-taking more naturally, allowing users to interrupt without errors. The mini version becomes the default in ChatGPT's Advanced Voice Mode, while paid subscribers can access the larger GPT-Live-1. Internally, the models connect to GPT-5.5 for search, reasoning, and agentic tasks, and can even present information visually. During a press briefing, OpenAI demonstrated the assistant staying silent for long periods to soak in context, then responding appropriately. Product lead Atty Eleti reported having 30- to 40-minute conversations while walking, highlighting the system's endurance for extended hands-free use.

The release positions voice as a future primary interface for complex work, with OpenAI noting over 150 million users already using ChatGPT's voice features. Competitors like Apple, Amazon, and startups like Monogram ($40M seed from DST/Lux) and Sesame are also pushing more expressive assistants. However, OpenAI emphasized that the new mode is not an AI companion—it includes safeguards for age-appropriate responses and resources for sensitive topics. Limitations remain: a live Hindi translation demo had a heavy American accent and unnatural phrasing. The models are optimized for "most spoken languages" but specifics were not given. As voice interfaces evolve, OpenAI's full-duplex approach could enable long-running agentic tasks, from research to coding, entirely through conversation.