OpenAI has overhauled its ChatGPT voice mode by introducing a new model called GPT-Live-1, designed to minimize unnecessary interruptions and wait for users to finish speaking if they pause mid-sentence. According to OpenAI research lead Kundan Kumar, GPT-Live-1 represents the company’s most intelligent voice model to date. The system automatically routes complex user queries to OpenAI’s advanced text models, such as GPT-5.5, whenever reasoning or web searches are required. The upgraded model can supplement spoken conversations with AI-generated visuals, including forecast charts and sports scores. The previous iteration relied on an older, turn-based architecture that frequently struggled with natural flow. The new framework introduces a full-duplex model capable of simultaneous listening and speaking. OpenAI product lead Atty Eleti explained that the architecture continuously and simultaneously processes an incoming stream of inputs while producing an outgoing stream of outputs, enabling real-time translation. Users can now instruct ChatGPT Voice to remain silent until explicitly called upon, and the AI utilizes conversational fillers such as “mhmm,” “yeah,” and “got it” to indicate active listening. OpenAI has implemented built-in guardrails to steer the model away from generating harmful content and to terminate sessions entirely during higher-risk interactions.