GPT-Live-1: full-duplex voice AI, and why Callaro was live with it on day one

On September 10, 2026, OpenAI launched GPT-Live-1 — a full-duplex voice model that can listen and speak at the same time, handle interruptions and background noise inside a single model, and delegate reasoning and tool use to a backend model of your choosing. Callaro has run GPT-Live-1 in production since launch day. This post explains what "full-duplex" actually changes for a phone call, why the delegation design matters, and how we turned a raw speech model into a complete voice agent — including the parts that are genuinely hard.

If you're new to how these systems fit together, start with the complete guide to AI voice agents. If you want the architecture underneath this post, read building on a frozen speech-to-speech model.

What "full-duplex" means — and why it matters on a phone call

Most voice AI to date has been half-duplex: the agent listens, then speaks, then listens again, taking turns like a walkie-talkie. It works, but it feels a beat slow, and interruptions are awkward — you have to stop the agent, wait for it to notice, and start over.

Full-duplex means the model listens while it speaks. That single change fixes the things that make AI calls feel robotic:

• Natural interruptions (barge-in). You can cut in mid-sentence and the agent hears you immediately and yields — the way people actually talk. (Barge-in is deceptively hard; more in our voicemail detection teardown, which shares the same audio-loop challenges.)
• Real backchannel. The agent can register an "mm-hm" or a correction without a full stop-and-restart.
• Lower perceived latency. Because listening never pauses, the conversation has no dead air between turns.
• Noise and overlap handled in-model. GPT-Live-1 handles interruptions and background noise within the model itself, rather than bolting on separate turn-detection logic.

For telephony specifically — where people talk over each other, background noise is constant, and even a half-second of dead air feels broken — full-duplex is the difference between "clearly a bot" and "wait, was that a person?"

The design that makes it usable: conversation vs. backend

Here's the part that matters for building real products. GPT-Live-1 is not one model doing everything. It splits the job:

• GPT-Live-1 owns the conversation — listening, speaking, turn-taking, interruptions, and the voice (12 voice options).
• A backend model of your choice owns the reasoning — deciding what to do, calling tools, looking things up, applying your business logic.

This is exactly the delegation pattern we'd already built our realtime stack around, which is why we were able to ship GPT-Live-1 the day it launched instead of re-architecting for it. The live model is the mouth and ears; a swappable delegate is the brain. Tool definitions and business logic live with the delegate, so they don't consume the live model's tight instruction budget, and you can pick the reasoning model independently of the voice. We go deep on this in the frozen speech-to-speech model post.

Why this makes for the best AI voice agent — with the right platform around it

A raw speech model, however good, is not a voice agent. GPT-Live-1 gives you a spectacular mouth and ears; everything that makes it do a job sits around it. What we layer on:

• In-call orchestration. Real actions mid-conversation — CRM writebacks, bookings, payment links, lookups — via tools, HTTP APIs and MCP servers, with a prior tool's output feeding the next. See orchestration.
• Any backend model. Because reasoning is delegated, you choose the brain — and you're insulated when models change or sunset (see model migrations and sunsets).
• Multilingual reach. Native-sounding languages and code-switching, which interact subtly with a speech model's own language coverage — see multilingual voice AI.
• Real-world audio. Noise, echo and background-media handling tuned for actual phone lines, including noisy markets and Indian PSTN — see noisy calls.
• Telephony + scale. Bring-your-own-carrier, bulk campaigns, and the pacing to run them.

Full-duplex is the leap in how it sounds; the platform is what makes it earn its keep.

The honest hard parts (that we'd already solved)

Shipping a speech-to-speech model in production surfaces problems demos never show. Because we'd been building on this pattern, we hit them early:

• Don't freeze the audio loop while the backend thinks. If you await the delegate inline, audio, transcripts and barge-in all stall. Delegation has to run as a background task.
• Every delegated token gets spoken — and billed as voice time. A long backend answer becomes minutes of speech. Answers must be kept tight.
• Send only the new turn to the delegate, not the whole transcript each time, or cost and latency balloon.
• Verbatim isn't guaranteed. A speech model voices text in its own way; for legally exact disclosures we fall back to a classic text-to-speech path.
• Sessions expire and have concurrency tiers. Long calls need reconnect handling, and bulk outbound has to respect per-tier session ceilings.

We wrote these up in detail in building on a frozen speech-to-speech model. The point isn't that GPT-Live-1 is hard — it's that a production platform has to have already answered these questions before it's safe to point at your customers.

Where this is going

GPT-Live-1 is a clear signal of the direction: the speech layer and the reasoning layer are separating, so the voice gets more human while the brain stays swappable and controllable. OpenAI is also moving its older realtime model line forward (GPT-Realtime is slated to retire in early 2027 in favor of a newer version), which is exactly why building on a stable delegation architecture — rather than hard-coding one model — is the durable choice. Our whole approach to model migrations is designed for this.

If you want to hear a full-duplex agent that also books, looks up, and follows through, see it on a real call. And if you're comparing platforms, our how to choose a voice AI agent checklist covers exactly what to test.