OpenAI Launches GPT‑Live‑1 API for Full-Duplex Voice Agents

OpenAI’s GPT-Live marks a shift from turn-based voice assistants toward continuous, full-duplex interaction, allowing the model to listen and speak at the same time. It can acknowledge users, pause, handle interruptions and delegate web searches or complex reasoning to GPT-5.5 while keeping the conversation moving. The technology matters because OpenAI is positioning voice as an interface not only for chat, but also for longer-running agent work. The company says more than 150 million people use ChatGPT Voice or Dictation each week.

OpenAI began rolling out GPT-Live globally on July 8, 2026, with GPT-Live-1 set as the default for Go, Plus and Pro subscribers and GPT-Live-1 mini for free users. On July 23, it extended voice to Work and Codex in the ChatGPT desktop app for macOS and Windows, letting users start, interrupt, redirect and coordinate multiple agents by speaking. Business customers receive one hour of Voice in Chat; extra use costs 5 credits a minute, while Voice in Work and Codex costs about 6 credits a minute. Standalone access is not offered on web or mobile.

Voice AI has traditionally either chained speech-to-text, an LLM and text-to-speech, or relied on a turn detector to decide when a user had finished. Both designs added delay and could mishandle pauses or background noise. OpenAI’s third-generation voice system, GPT-Live, instead uses a full-duplex model that listens and speaks simultaneously, while deeper reasoning runs outside the live audio path. The separation is key to making ChatGPT Voice feel more conversational without sacrificing intelligence or scale.

On Aug. 3, 2026, OpenAI detailed a six-month rebuild spanning inference, context management and media transport; the company did not disclose development spending. Rewriting the media frontend and inference logic in Go from Python asyncio lifted p95 frame-delivery smoothness to the previous system’s p50 level. WARP cuts media and data startup from six network round trips to one, while complex requests are delegated asynchronously to frontier models including GPT-5.5. The architecture underpins the ChatGPT Voice update launched on July 8.

OpenAI continued to lead the technology industry in 2026, with its valuation reaching $150 billion. Earlier voice assistants relied on half-duplex architectures that caused noticeable delays in human-machine conversations. GPT-Live’s full-duplex capability is designed to overcome those constraints and usher AI voice systems into an era of real-time, simultaneous conversation. The development could prove significant for the global voice customer-service market, which is worth tens of billions of dollars.

OpenAI formally launched the GPT-Live-1 voice model and a mini version on July 16, 2026, initially making them available to Plus subscribers paying $20 a month. The models can listen and speak simultaneously, interject at appropriate moments and deliver a more natural conversational experience. They separate voice interaction from the reasoning layer, delegating complex tasks to GPT-5.5 running in the background. In an official demonstration, the AI simultaneously handled flight bookings, weather queries and itinerary planning.