GPT-Live-1 and its smaller sibling replace the previous voice stack with something architecturally distinct: full-duplex operation, where the model is continuously deciding — many times per second — whether to speak, hold, interrupt, or invoke a tool. That's a different problem from the pipeline approach of listen-then-transcribe-then-respond, and it shows up in turn-taking behavior: acknowledgment cues like "mhmm" and the ability to handle silence without treating it as a prompt boundary. GPT-Live-1 mini is now the ChatGPT default; paid subscribers get the larger model. The delegation pattern is worth noting. When a conversation demands web search, reasoning, or heavy-lift tasks, GPT-Live-1 hands off to GPT-5.5 behind the scenes. That makes it less of a monolithic voice model and more of a real-time coordination layer — the voice surface stays snappy while the cognitive heavy lifting goes elsewhere. Both models are available via the OpenAI API, so the same architecture is open to third-party apps.