OpenAI has launched GPT-Live, a new language model for ChatGPT that enables real-time listening and speaking during conversations. The model uses a full-duplex architecture, allowing it to respond to interruptions and use filler words like 'mhmm' to maintain natural conversation flow. Two versions of GPT-Live are now available: one for paying users and one for free users. Both versions support iOS, Android, and ChatGPT.com, with API access planned for the future. The update also includes visual cards for real-time information such as weather and stock prices, along with revamped voices for the model. Source: thedecoder
GPT-Live delegates complex tasks like web searches and logical reasoning to the background GPT-5.5 model, improving answer quality while maintaining the conversation. This architecture allows GPT-Live to stay connected to the latest frontier models at all times. Users can choose reasoning levels from 'Instant' to 'High' for different task complexities. The model also handles interruptions, pauses, and allows users to slow down the conversation. OpenAI claims this delegation closes a major weakness of earlier live models, which struggled with tasks requiring external knowledge. Source: thedecoder
Benchmark results show significant improvements with GPT-Live. On the GPQA test for scientific reasoning, GPT-Live-1 achieved 84.2% accuracy at the high reasoning level, compared to 45.3% in Advanced Voice Mode. On the BrowseComp benchmark for web search tasks, GPT-Live-1 scored 75.2%, while Advanced Voice Mode managed only 0.7%. The Mini version reached 31.6% accuracy. Additionally, GPT-Live outperformed Advanced Voice Mode on the tau3 Voice Telecom benchmark, solving about 65% of support tasks in 385 seconds compared to 30% for Advanced Voice Mode. Source: thedecoder
