OpenAI’s voice layer makes interruptions part of the interface
Voice agents usually force a choice: listen or talk. GPT-Live-1 does both, handling interruptions and turn-taking in one stream while a separate model handles deeper reasoning and tools. That split costs $0.05 per minute for the voice layer, before backend fees, and includes 12 voices. For support, tutoring, and phone workflows, the meaningful change is less awkward silence and fewer brittle speech handoffs.