동시에 듣고 말하는 오픈AI 'GPT-Live' 등장

ChatGPT can now listen and talk at the same time, making AI conversations seem more human Matthias Bastian View the LinkedIn Profile of Matthias Bastian Jul 8, 2026 Key Points OpenAI has introduced GPT-Live, a new language model for live conversations in ChatGPT that can listen and speak simultaneously thanks to a full-duplex architecture. The model responds to interruptions, uses filler words like "mhmm" to keep the conversation flowing naturally, and is available in two versions: one for paying users and one for free users. For complex tasks that require web searches or logical reasoning, GPT-Live delegates the requests to the GPT-5.5 model in the background while maintaining the conversation, significantly improving answer quality. Ask about this article… Search OpenAI introduces GPT-Live, a new generation of voice models with full-duplex architecture. The model can listen and speak at the same time and hands off complex tasks to GPT-5.5 in the background. With GPT-Live, OpenAI is shipping a new class of AI voice models designed to make talking to ChatGPT feel more like a real conversation. Two versions are rolling out worldwide right away. GPT-Live-1 is for paying users on Go, Plus, and Pro plans, while GPT-Live-1 mini is available on free accounts. Both work on iOS, Android, and ChatGPT.com . OpenAI plans to add API access soon, and developers can sign up through a form . Instead of the old rigid back-and-forth where users speak and the AI responds, GPT-Live uses a full-duplex architecture. The model can listen and talk at the same time, much like a real conversation between people. Earlier this year, Nvidia released a similar open-source model called PersonaPlex . Ad OpenAI says GPT-Live makes decisions multiple times per second about whether to speak, keep listening, pause, interrupt, or call a tool. The model can use filler phrases like "mhmm" or "got it" to signal it's following along. Users can interrupt, take a moment to think, or ask the model to slow down. Ad DEC_D_Incontent-1 Compared to the previous "Advanced Voice Mode," people preferred GPT-Live-1 in 75.7 percent of cases and GPT-Live-1 mini in 69.2 percent, OpenAI says. Beyond the new conversational features, the update adds visual cards that ChatGPT Voice can show during a conversation for things like weather, stock prices, or sports scores. OpenAI also revamped the nine available voices for GPT-Live. At launch, GPT-Live doesn't support Voice with video or screen sharing in ChatGPT, though OpenAI says those features are coming soon. The older Standard and Advanced Voice Mode with these features will stick around for now. Ad Background delegation to GPT-5.5 closes the intelligence gap The biggest change is the split between live conversation management and actual reasoning. When a question needs a web search, reasoning, or agent-like capabilities, GPT-Live hands it off to a background model. Right now, that's GPT-5.5. While the background model works, GPT-Live keeps the conversation going. OpenAI says the architecture is built so GPT-Live stays connected to the latest frontier models at all times. Users can also pick a reasoning level, from "Instant" for quick answers to "Medium" and "High" for tasks where ChatGPT should spend more time thinking. Ad DEC_D_Incontent-2 This delegation fixes a major weakness of earlier live models, which answered questions based only on their own capabilities. That made them fall far behind frontier models and left them barely usable for serious work. Ad With GPT-Live, that gap appears to be closing, and benchmark scores for knowledge are drastically better. On the GPQA test for scientific reasoning, GPT-Live-1 hits 84.2 percent accuracy at the high reasoning level, compared to 45.3 percent in Advanced Voice Mode. The gap is even wider on BrowseComp, a benchmark for agent-based web search. GPT-Live-1 scores 75.2 percent. Advanced Voice Mode manages 0.7 percent. GPT-Live also leads on OpenAI's internal tau3 Voice Telecom test, which evaluates full-duplex voice agents on realistic telecom support tasks. Depending on reasoning level, GPT-Live-1 completes tasks at a much higher rate while also finishing faster. Advanced Voice Mode does worst, with the lowest success rate and the longest processing time. More human-sounding AI raises safety stakes People have a well-documented tendency to treat AI models like humans and let themselves be talked into all sorts of actions , sometimes with serious consequences . Research shows that heavy voice model users are even more prone to emotional dependency on AI . An AI system that sounds and responds even more like a person will likely amplify that risk. OpenAI says it has built safety measures that kick in even while the user is still speaking. The system can steer the model toward safer responses, show extra safety information, or end the conversation entirely in high-risk situations. Crisis hotlines appear when topics related to self-harm come up. For younger users, OpenAI trained the model to behave in an age-appropriate way. Parents can use parental controls to decide whether their child can access ChatGPT Voice and will get notified in high-risk situations. GPT-Live also isn't designed to mimic real voices and only uses predefined ones. Full details on the safety measures are available in the System Card . AI News Without the Hype – Curated by Humans Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section. Subscribe now Source: OpenAI