OpenAI2026-08-04 06:41:33OpenAI details GPT-Live voice system with full-duplex replies and background model routingOpenAI has shared new architectural details about GPT-Live, its latest voice system, focusing on a capability that older voice assistants often struggled to deliver: speaking while listening. According to the company, GPT-Live uses a full-duplex design that allows it to insert short acknowledgements such as “mm-hmm” or “okay” while a user is still talking, pick up the conversation during natural pauses, and recover more gracefully when interrupted instead of breaking the exchange. To make real-time voice work at ChatGPT scale, OpenAI said it rebuilt the voice stack from the client side through the model layer. The system sends live speech interaction through a dedicated fast path for immediate responses, while more demanding jobs such as web search, deeper reasoning, or agent actions are handed off asynchronously to a stronger background frontier model. At the time described by OpenAI, that background model was GPT-5.5, which the company said launched in April. This setup lets the main model keep talking rather than pausing for a result. OpenAI also said it sharply reduced voice session startup latency. In human blind tests lasting 5 to 10 minutes, the company reported that GPT-Live-1 achieved about a 75.7% preference rate over the original Advanced Voice Mode, while GPT-Live-1 mini posted about 69.2%. GPT-Live itself went live in July; the latest disclosure explains how the system works behind the scenes.2050
OpenAI2026-07-16 07:29:42OpenAI shows GPT-Live handling flights, weather and itinerary tasks in one voice conversationOpenAI on July 15 released a demo showing GPT-Live carrying out multiple tasks in the background while keeping a single voice conversation going. In the video, the model checks flights, pulls local weather and drafts an itinerary in real time. OpenAI said the key change is that the system does not need to wait for a user to finish an entire turn before responding. GPT-Live went live on July 8 and replaced ChatGPT’s previous Advanced Voice Mode. The company describes the system as full-duplex, meaning it can listen to a user and generate spoken output at the same time. During a conversation, it can use short acknowledgments such as “mhmm” and “yeah,” or stay quiet when the user is thinking. OpenAI said GPT-Live comes in two versions: GPT-Live-1 as the default for paid plans and GPT-Live-1-mini for free users. When a conversation needs deeper reasoning or search, GPT-Live can hand that task to GPT-5.5 while continuing to manage the live exchange. The company also said GPT-Live still does not support screen sharing or camera input, which remain available only through the older ChatGPT Voice Mode. In tests lasting five to 10 minutes, both GPT-Live versions outperformed the previous Advanced Voice Mode across conversation preference, turn-taking, interruption handling, fluency and naturalness.2090
AI2026-07-10 06:39:13Ex-OpenAI CTO's Thinking Machines Lab Unveils Full-Duplex AI for Natural Voice ConversationsFormer OpenAI CTO Mira Murati’s Thinking Machines Lab releases a research preview of Interactive Models featuring real-time full-duplex voice AI with approximately 0.40-second response time, aiming to make voice interactions as natural as phone calls.710
