OpenAI Launches Third-Generation Voice Model GPT-Live, Supporting Full-Duplex Real-Time Conversations
OpenAI has rolled out GPT-Live to ChatGPT users worldwide, with free users accessing GPT-Live-1 mini and paid users defaulting to GPT-Live-1. It separates complex reasoning from real-time conversation, allowing users to listen while speaking.
Using a full-duplex architecture, it indicates listening with filler words, enabling quick interactions or quiet waiting; the backend delegates search reasoning to GPT-5.5, presenting weather and stock market visual cards.
Source: Public Information
ABAB AI Insight
OpenAI's previous voice models have iterated through multiple generations, and this time, the GPT-Live full-duplex architecture continues its progress from walkie-talkie style to natural conversation, optimizing real-time performance by separating reasoning tasks.
On the capital path, the free/paid tier drives upgrades, with resources tilted towards the backend GPT-5.5, motivated by enhancing user stickiness, and strategically expanding daily usage scenarios of ChatGPT through visual aids and natural interactions.
Similar to Google Gemini's real-time voice updates and other large model voice races, OpenAI is currently in a leading phase of transforming voice interaction from functionality to human-like natural conversation.
Essentially, this represents a technological substitution: separating full-duplex and backend reasoning, with the mechanism stemming from architectural optimizations that break through dialogue delays, shifting pricing power from traditional voice assistants to multimodal real-time AI platforms and reshaping human-computer interaction paradigms.
ABAB News · Cognitive Laws
Full-duplex means natural conversation, and AI's listening ability defines interaction quality.
Separating reasoning backend from real-time frontend ensures complex tasks do not interrupt a smooth experience.
Visual cards assist understanding, accelerating the evolution of voice models for daily AI substitution.