OpenAI's GPT-Live-1 learns to wait through pauses, cuts interruptions
OpenAI's new voice model detects mid-sentence pauses and avoids jumping in, rolling out to Plus and Team subscribers today.
OpenAI released GPT-Live-1 this week, a voice model designed to handle conversational pauses without interrupting. The model waits when a user trails off mid-sentence instead of treating every silence as a turn-taking cue. Research lead Kundan Kumar described it as "talking to another person" during a press briefing. Unlike post-processing filters that rely on silence thresholds or external timers, GPT-Live-1 learns conversational rhythm as a native capability baked into training.
GPT-Live-1 is rolling out to ChatGPT Plus and Team subscribers starting today, with free-tier access coming in the coming weeks. The update addresses user complaints about Advanced Voice Mode cutting speakers off mid-thought—a common scenario where users pause to recall a name or date, only to have the model assume the turn had ended and start responding. Google's Gemini Live and Anthropic's voice-capable Claude models have both shipped pause-aware features in recent months, making conversational flow a competitive front across the industry.
