OpenAI's new ChatGPT voice model speaks and listens simultaneously, interrupts less

OpenAI has unveiled a revamped voice experience for ChatGPT powered by GPT-Live-1, a model designed to feel closer to talking with another person. The new system processes speech input and generates responses simultaneously—what OpenAI calls "full duplex" capability—enabling more fluid conversations with fewer awkward interruptions.

How it works#

The GPT-Live-1 model can speak and listen in parallel, fundamentally changing how voice interactions unfold. Rather than waiting for you to finish speaking before responding, it now understands context dynamically, allowing natural pauses without losing the thread. If you hesitate mid-sentence, the assistant will wait rather than jumping in prematurely.

The model also integrates with OpenAI's more powerful text-based models, like GPT-4o, delegating complex reasoning and web searches when needed. This handoff happens quickly, so users get research results discussed aloud almost immediately rather than after a noticeable delay.

New capabilities#

Real-time translation is now possible—ChatGPT can translate your words as you speak instead of waiting for you to stop. Users can also request the assistant to stay silent until prompted, with acknowledgments like "mhmm" or "got it" replacing intrusive interruptions.

Weather, stocks, and sports discussions now include AI-generated visuals displaying relevant data such as scores or forecasts, adding context without requiring follow-up questions.

Safety and access#

OpenAI built safeguards to prevent harmful outputs and includes "expert-vetted crisis helpline support" when users discuss self-harm. The model adapts responses for younger users marked as teens.

The rollout covers iOS, Android, and web browsers. ChatGPT's paid tiers—Go, Plus, and Pro—access GPT-Live-1, while a leaner GPT-Live-mini-1 variant serves free users.

Read OpenAI's full announcement.