OpenAI's GPT Live goes full-duplex; the developer API is next
The architecture behind ChatGPT's voice, in OpenAI's own words: one-round-trip startup, live interruptions, and reasoning handed off to GPT-5.5.
Copy markdown
Full-duplex, not turn-based
GPT Live listens and talks at the same time — it backchannels (“mhmm”), stops mid-sentence when you cut in, and stays quiet while it thinks, re-deciding many times a second whether to speak, listen, or call a tool. This is the conversational bar your own voice agents now get judged against.
Startup: six round trips down to one
A new transport OpenAI calls WARP collapses session setup from six network round trips to one, putting first audio under a second. In voice, that opening delay is the whole first impression — and the hardest thing to paper over later.
Hard reasoning gets offloaded to GPT-5.5
GPT Live runs the conversation layer itself and hands web search, multi-step reasoning, and agentic work to GPT-5.5 Instant or Thinking, folding the result back without breaking the flow. The pattern worth copying: a cheap always-on voice model, frontier reasoning only when a turn needs it.
The developer API is next, not now
GPT Live is ChatGPT-only at launch; OpenAI says a GPT-Live API (full GPT-Live-1 plus a cheaper mini) is coming — reportedly a developer waitlist that’s “weeks, not months” out. Until it lands, voice agents still ship on the Realtime API (gpt-realtime-2.1), at about $32/$64 per million audio tokens in/out.
What it unlocks — and what to budget
Full-duplex plus single-pass live translation makes real-time interpreters, interruptible phone agents, and always-listening copilots practical. Early estimates peg GPT-Live API pricing at 3–5× today’s text APIs, so design around audio-minute costs from day one.