OpenAI released GPT-Live-1 today. Less than 24 hours later, any Glytos voice agent can run on it. No waitlist and no private beta: open an agent, switch it over, and place a call.
What GPT-Live is
Voice agents have mostly been built as a relay. Speech goes to a transcriber, text goes to a language model, the reply goes to a voice. Three services, three handoffs, and the caller waits out every one of them.
GPT-Live collapses that into a single model that listens and speaks at the same time. It is full duplex, which is a stronger claim than "you can interrupt it": the model is still hearing you while it talks, so adding a detail halfway through its answer is an ordinary thing to do rather than a race against its turn.
It also splits thinking from talking. The conversation stays with the live model while reasoning and tool calls are handed to a separate backend model, so the line does not go quiet while something is looked up.
What it is like on a real call
We called one from a phone. The back-and-forth feels natural in the way the demos suggest: you can cut in, add something you just remembered, or change direction without waiting for a sentence to finish, and the model picks it up rather than talking over you.
That is an impression from a handful of calls rather than a benchmark. We have not timed it, and we would rather say nothing than publish a number we have not taken.
Three things that stop being true
This is the part worth reading before you build on it, whoever you build with. A full-duplex model quietly invalidates a few assumptions that every voice stack currently holds, and none of them fail loudly.
Muting the caller is no longer a defence
In a cascaded stack it is routine to mute the caller's microphone while the agent is speaking. It is the standard guard against the agent's own audio coming back through a speaker and being transcribed as a new caller turn, and against a half-finished sentence being cut off by a cough.
On a full-duplex model that same mute does not prevent an interruption. It stops the model hearing anything at all for the whole of its turn, which removes the one capability it was chosen for. The model is meant to own that decision: it hears the interruption and decides whether to yield.
Nothing errors when you get this wrong. The call connects, the agent talks, and it simply never notices anyone talking back.
A provider name no longer implies a behaviour
OpenAI now ships two speech engines under one name, reached through two different API surfaces, and they differ in ways that matter. The clearest is scripted speech: the realtime models accept a per-response instruction that makes the model say an exact line, and GPT-Live does not, because its instructions belong to the session.
If your agent has a closing line it must say word for word, or a fixed prompt it falls back to when it cannot understand the caller, the mechanism that delivers it is not the same on both. Code that decides behaviour from the provider name will do the wrong thing on one of them, and the symptom is a line that is never spoken.
A voice agent's cost now has two moving parts
OpenAI bills GPT-Live audio by the second and the backend model separately on its own tokens. Both figures are theirs and published, and the point here is only the shape: a conversation that used to be priced as one thing is now priced as two, and they move independently. A chatty caller and a caller who asks one hard question cost differently, in different directions, than they used to.
Worth knowing before you model unit economics on a per-minute number alone.
Trying it
On any voice agent, open the model settings, switch the speech engine to realtime and pick GPT-Live as the model. Choose a voice and call it. Everything else the agent already has keeps working: its prompt, its knowledge base, its tools, its phone number.
Early days
It has been live here for hours rather than weeks, so this is a first look and not a verdict. We like what we have heard so far. The measurements that would turn that into a claim come later, and when they do they get their own post.