OpenAI launches GPT-Live voice models
GPT-Live-1 and GPT-Live-1 mini replace Advanced Voice Mode in ChatGPT with full-duplex audio and real-time background task delegation to GPT-5.5.
OpenAI released two new voice models on July 8: GPT-Live-1 and GPT-Live-1 mini. Both are rolling out globally on iOS and Android and replace Advanced Voice Mode as the default ChatGPT voice experience.
GPT-Live-1 is the default for Plus, Pro, and Go subscribers. GPT-Live-1 mini is the default for Free users. API access for both models is listed as coming soon; a developer waitlist is open.
Full-duplex audio
The defining change is full-duplex processing. Unlike prior voice models that wait for you to finish speaking before generating a response, GPT-Live processes input and output simultaneously. The models deliver back-channel cues mid-sentence, such as brief acknowledgments while the caller is still speaking, which matches conversational turn-taking patterns more closely than the previous push-to-talk model.
Both models outperform Advanced Voice Mode on GPQA, BrowseComp, and OpenAI’s internal tau-3 Voice Telecom benchmark in head-to-head evaluations.
Background delegation
The second major change is background delegation. GPT-Live models can route web search, multi-step reasoning, and agentic tasks to GPT-5.5 while the voice conversation continues uninterrupted. The caller hears a continuous conversation; the model executes a task in a parallel thread and surfaces the result when ready.
OpenAI cited over 150 million weekly ChatGPT voice and dictation users at launch.
What this means for builders
For end users, the transition from Advanced Voice Mode is automatic today.
For developers building voice agents, background delegation shifts the design constraint. A prior design assumption was that the model had to pause or stall while looking up account data or running a calculation. That constraint is lifted: the live conversation can continue while a GPT-5.5 agent queries a backend and returns a result. The limit becomes backend latency and the model’s ability to hold the caller’s attention, not the voice model itself.
The waitlist for API access is the key action for anyone building now. The realtime-2.1 model update from July 6 improved API-tier latency by 25%; GPT-Live is the consumer-layer equivalent built on the same trajectory. The guide on AI in customer support covers where voice agents fit in a broader support stack.