OpenAI is launching two new voice models, GPT-Live-1 and GPT-Live-1 mini, significantly enhancing ChatGPT's voice capabilities. Unlike the previous turn-based model that required users to stop speaking before responding, the new duplex architecture allows simultaneous input processing and output generation. This enables more natural, real-time conversations, the ability to interrupt, and even live translations.

The GPT-Live models can also delegate complex tasks, like web searches or reasoning, to other OpenAI models such as GPT-5.5 in the background. Users can explicitly ask the models to slow down, and they provide verbal acknowledgements like "mhmm" while listening. The new voice mode also improves focus amidst background noise and can generate visual cards for topics like weather and stocks.

Enhanced safety features are included, with the system capable of steering conversations toward safer responses or ending them in high-risk situations. Parental controls allow disabling the feature for teens, with notifications sent to guardians if a conversation veers toward self-harm. GPT-Live is rolling out today to ChatGPT on Android, iOS, and the web, with GPT-Live-1 available for paid subscribers and GPT-Live-1 mini for free users.