OpenAI launches GPT-Live full-duplex voice models with 25% latency cut

OpenAI's GPT-Live is a bet that voice becomes the primary interface for consumer and developer AI. GPT-Live-1 and its mini variant are full-duplex—they listen and speak at the same time, enabling barge-in interruptions and near-simultaneous translation rather than the turn-based cadence of the old Advanced Voice Mode. The models roll into ChatGPT immediately on iOS, Android and web, and OpenAI says API and Codex support are in progress. On the developer side, gpt-realtime-2.1 and gpt-realtime-2.1-mini in the Realtime API cut p95 latency by at least 25%, the metric that most determines whether a voice app feels conversational.

Mechanically, full-duplex audio requires the model to process incoming audio while generating output, a harder engineering problem than the half-duplex pipelines most assistants use. Greg Brockman called it 'intelligent voice AI that feels like having a natural conversation,' adding the team feels it is 'still just scratching the surface.'

Competitively this targets Google's Gemini Live, which reviewers still rate as the cheaper option, and Apple's revamped Siri. Community reaction on The Verge and developer threads was split: one tester called GPT-Live 'smarter but more robotic' than the old GPT-4o voice, and developers flagged that realtime-tier pricing breaks consumer-product unit economics. Watch whether OpenAI closes the price gap with Gemini Live and how quickly API access lands, since that determines third-party adoption.