OpenAI introduced GPT-Live on Thursday, a voice model built for full-duplex conversation: it can listen and speak at the same time instead of waiting for a speaker to finish before responding. The company says the model handles interruptions, backchannels such as “mm-hmm,” and natural turn-taking, and can hand off demanding reasoning to GPT-5.5 mid-conversation without stalling the exchange.
That distinction matters more than it sounds. Most voice assistants, including earlier ChatGPT voice mode, run a turn-based pipeline: transcribe the user’s speech, wait for a pause, generate a response, then speak it. Full-duplex systems process audio continuously in both directions, so the model can react while the user is still talking, closer to how a real conversation partner behaves.
GPT-Live is a separate product from GPT-Realtime-2.1-mini, OpenAI’s existing low-latency voice API. Where Realtime models optimize for fast turn completion inside a request-response cycle, GPT-Live is built around simultaneous audio streams and a harder problem: deciding, in real time, when to interrupt, when to yield, and when a question is complex enough to route elsewhere.
That routing is the second half of the announcement. According to OpenAI’s blog post, GPT-Live can delegate demanding tasks to GPT-5.5 while keeping the spoken exchange going, so the caller hears a natural pause or acknowledgment rather than dead air while the heavier model works. This is the engineering problem voice-agent builders have wrestled with since real-time APIs shipped: reasoning models are too slow for live conversation, and fast models are too shallow for hard tasks. Splitting the two, with a fast model managing the channel and a slow model handling substance, is the same architectural bet behind Google’s Gemini Live and its underlying Project Astra research.
OpenAI’s post does not include independent benchmark results, latency figures, or detail on how GPT-Live decides when a task warrants a GPT-5.5 handoff. That decision boundary is where full-duplex systems tend to break down in practice. Cut over too eagerly and the assistant talks over itself. Cut over too late and the illusion of a live conversation collapses anyway, since the whole pitch of full-duplex voice is that the seams do not show.
The competitive stakes are direct. Voice-first products, from customer service bots to in-car assistants, have been bottlenecked by the same tradeoff GPT-Live claims to solve. If the model holds up outside a curated demo, it pressures every voice API vendor still shipping half-duplex pipelines to explain why a hold-and-respond design is still good enough for 2026 users.
Teams currently building voice agents on GPT-Realtime-2.1-mini should test GPT-Live against real interruption-heavy call transcripts before treating it as a drop-in upgrade. Full-duplex behavior changes the entire conversation design, not just the model underneath it.
Reported by OpenAI on its company blog (openai.com), July 9, 2026.
