OpenAI announced the launch of GPT-Live, a new generation of voice models designed to make conversations with AI feel more natural, on July 8, 2026. The technology powers ChatGPT Voice and enables real-time, two-way communication by allowing the model to listen and speak simultaneously. GPT-Live can respond with phrases like 'mhmm' or 'yeah' to show it is attentive, engage in quick back-and-forth, or remain silent when needed. This results in a more fluid and realistic interaction experience. The model also delegates complex tasks to the latest frontier models, ensuring the conversation continues uninterrupted while background work is handled. GPT-Live will initially use GPT-5.5 in the background and will be updated as newer models are released. It is being rolled out globally to ChatGPT users and will soon be available via API for developers and enterprises.
GPT-Live represents a significant advancement over previous voice models by using a full-duplex architecture, allowing continuous interaction and real-time responses. This approach enables the model to make frequent decisions on whether to speak, listen, or pause, creating a more natural flow. Additionally, GPT-Live can offload tasks requiring search, reasoning, or complex execution to other models like GPT-5.5, maintaining conversation while handling background work. The model also supports varying levels of reasoning, from fast responses using GPT-5.5 Instant to more detailed answers with Medium and High reasoning modes. These improvements are expected to enhance the usability of voice-based interactions for everyday tasks.
OpenAI emphasized that GPT-Live is part of a broader effort to enable more natural human-AI interaction, where collaboration feels as fluid as working with another person. The company highlighted that previous voice systems, such as cascaded models or turn-based systems, had limitations in responsiveness and flow. GPT-Live addresses these issues through architectural changes, allowing for continuous interaction and seamless integration of advanced capabilities. The new model is also being tested for safety, with expanded evaluations to ensure it meets the standards for real-world voice interactions.
Source: openai
