OpenAI upgrades ChatGPT voice with real-time, full-duplex models

The new models replace the existing Advanced Voice Mode in ChatGPT, with the mini version serving as the default and the larger GPT-Live-1 reserved for paid subscribers. Unlike previous iterations that relied on a sequential chain of speech-to-text, LLM processing, and text-to-speech, the updated architecture processes queries directly through advanced models like GPT-5.5. This integration enables the assistant to handle reasoning, search tasks, and visual data presentation while maintaining a continuous audio stream.

Product lead Atty Eleti noted that the technology is designed to sustain long-form engagement, citing personal experience with 30- to 40-minute sessions. Despite these improvements, the system remains a work in progress; demonstrations of Hindi translation revealed a persistent American accent and a rigid, bookish tone. While OpenAI is positioning voice as a future pillar of computing, it maintains that the tool is not intended as an AI companion, embedding specific safeguards to manage sensitive topics and provide age-appropriate responses for younger users.