OpenAI's GPT-Live-1 Makes Voice Full-Duplex and Delegates Hard Reasoning to GPT-5.5

OpenAI·high signal

GPT-Live-1 and GPT-Live-1 mini replace turn-based voice with a full-duplex architecture that processes input while generating output, deciding many times per second whether to speak, keep listening, pause, interrupt, or call a tool. The architecturally interesting part for builders is the delegation pattern: when a query needs web search or deeper reasoning, GPT-Live hands off to the frontier model in the background and folds the result back into the conversation while continuing to talk. GPT-Live-1 becomes the default voice model for Go, Plus, and Pro tiers; mini serves free users. OpenAI is explicitly positioning voice as the primary interface for supervising long-running agentic work.