GPT-Live and the Full-Duplex Unlock Nobody Is Explaining

On July 8, 2026, OpenAI replaced Advanced Voice Mode in ChatGPT with a new family of models called GPT-Live.^[1] Two of them: GPT-Live-1 for Go, Plus, and Pro users, and GPT-Live-1 mini for the free tier.^[3] By the end of the week the paid rollout was complete on iOS, Android, and the web.^[11]The coverage was almost uniformly about vibes. It talks while you talk. It says “mhmm.” It feels human.

All true, and all beside the point. The word buried in every one of those articles is full-duplex, and it is doing more work than any other word in the launch. Full-duplex is a term borrowed from telecom: a channel where both ends can transmit at the same time, like a phone line. Half-duplex is a walkie-talkie. One side talks, then releases the button, then the other side talks. Every mainstream voice assistant before this one, Siri, Alexa, and yes, Advanced Voice Mode, was a walkie-talkie wearing a personality.

Making it a phone line instead is not a UX polish job. It is a rewrite of the entire inference loop. That is the story.

What half-duplex actually cost you

The old architecture was a pipeline, and you can draw it on a napkin. Your microphone streams audio into a voice activity detector. The detector waits for you to stop talking. That decision is called endpointing (the system guessing that your utterance has ended), and in practice it is a silence timer. Once the timer fires, an automatic speech recognition model turns your audio into text. The text goes to a language model, which generates a full response. The response goes to a text-to-speech model, which renders audio. Then the audio plays.

Four models, in series, each one waiting on the one before it. Every stage adds delay and every stage adds its own errors on top of the previous stage's errors. One analysis of that stacked pipeline put the delay before the first word of a response at up to roughly 1,700 milliseconds.^[9] Nearly two seconds. Go count that out loud in a conversation with a person and watch how weird it feels.

And here is the part that annoyed me most as an engineer. The endpointing timer is a hard tradeoff with no good setting. Make it short and the assistant cuts you off mid-thought every time you pause to think. Make it long and every exchange feels like satellite delay. There is no value that works, because the timer is trying to solve a problem that shouldn't exist. Humans don't wait for silence to decide you're done. We predict it, from prosody, syntax, and content, while you are still talking.

The number that defines the whole problem is 230 milliseconds

Conversation analysis has a well-known result: the gap between one human finishing a turn and the next one starting is roughly 200 milliseconds. The Moshi paper out of Kyutai, the first real-time full-duplex spoken language model and still the clearest public writeup of the architecture, anchors on 230 ms as the measured human-to-human figure and reports a theoretical latency of 160 ms and about 200 ms in practice.^[7]

That number is brutal, because 200 ms is shorter than it takes most people to plana sentence. Which means humans are not responding to the end of your turn. We are composing our response while you are still speaking, and firing it the instant we predict you're done. A voice system that only starts thinking after you stop has already lost. It is structurally incapable of hitting the target.

So the full-duplex bet is: stop treating speech as a request-response API and start treating it as two continuous token streams, yours and mine, running at the same time. Moshi does this literally, modeling the user's audio stream and its own audio stream in parallel, with interruptions, overlaps, and backchannels emerging from the token stream instead of being coded as explicit turn boundaries.^[7] GPT-Live is the first time a system built on that idea has shipped to a few hundred million consumers.

The four hard parts

If you have ever tried to build this, you know the demo is easy and the product is not. Four problems eat the schedule.

Echo cancellation.The model is speaking through a speaker, and its own microphone is picking that up. If you don't subtract your own output from your own input, the model hears itself, decides someone is talking, and interrupts itself. Acoustic echo cancellation is a solved problem in telephony, but it gets much harder when the far-end signal is being generated in real time and the device is a phone on a table in a car. Every full-duplex system lives or dies on this and nobody writes about it.

Barge-in. Interrupting. When you start talking over the model, it needs to shut up fast, and it needs to know what it actually managed to say before it got cut off. That last part is subtle. If the model generated three sentences but the audio only played one and a half, its own context has to be truncated to what the user actually heard, not what it planned. Get this wrong and the model confidently references information it never delivered.

Endpointing, still.Full-duplex doesn't delete the problem, it converts it from a timer into a continuous prediction. The model is now guessing, at every frame, whether this pause is a breath or a finished thought. That's why GPT-Live can “stay quiet when you need a moment to think” instead of pouncing on the first 500 ms of silence.^[1]

Streaming inference under a latency budget.You can't batch. You can't wait for the full input. Every component has to emit partial output while consuming partial input, and the whole chain has to fit inside a couple hundred milliseconds on infrastructure serving millions of concurrent sessions. This is a systems problem as much as a modeling one, and it is the reason voice stayed bad for so long even after the language models got good.

The genuinely clever bit: they split the model in two

Here is the design decision I think is the most underrated thing in the launch. A full-duplex model has to be small and fast enough to make speak-or-listen decisions many times a second. A model that good at physics questions has to be big and slow. Those requirements are in direct conflict, and for two years that conflict was the reason voice assistants were dumber than the chat window on the same website.

OpenAI resolved it by decoupling the layers. GPT-Live is the conversational front-end. When a question is easy, it answers directly. When the question needs web search, deep reasoning, or agentic work, it hands the task off to a frontier model running in the background, keeps the conversation alive while that runs, and folds the result back in when it lands. At launch the background model is GPT-5.5.^[1,4] Because the two layers are separate, the reasoning model can be swapped later without retraining the voice.^[4]

Now the money question

Voice is the interface most people already know how to use. It needs no onboarding, no app, no literacy in English or in software. The reason it never mattered commercially is that it was bad. Not slightly bad. Bad in the specific way that makes people hang up.

Look at drive-throughs, which is the most honest test bench in the world: background noise, accents, kids in the back, and someone changing their order halfway through a sentence. McDonald's ran a multi-year IBM pilot at roughly 100 locations and pulled the plug in June 2024 with order accuracy around 85%, short of the 90 to 95% bar it wanted.^[14]Wendy's FreshAI, built with Google, reports filling orders correctly about 86% of the time without an employee stepping in.^[14] Taco Bell has been the aggressive one and expanded voice AI to nearly 900 US restaurants by July 2026.^[13] Across the category, roughly a fifth of AI-assisted drive-through orders still need a human to intervene, and accuracy only reaches 95% when they do.^[14]

Every one of those failures is a half-duplex failure. The customer says “a number three, no wait, make it a number four,” and a turn-based state machine has no idea what to do with that. A model that listens while it speaks does. That doesn't mean GPT-Live solves drive-throughs tomorrow, and there is no API for it yet.^[6] It means the architectural excuse just expired.

Scale it up. The global call center market was around $364 billion in 2025.^[16] The Philippines alone runs roughly 16% of global outsourcing on the back of a workforce measured in the millions.^[16] That is not a software market being disrupted. That is a labor market. And the thing that has protected it is not that the work is hard, it is that machines could not hold a conversation. Real-time translation is the same story from a different angle: OpenAI already sells a translation model in the Realtime API at $0.034 per minute covering 70-plus input languages.^[15] Two cents a minute against a human interpreter is not a competitive market, it is a different universe.

What OpenAI is quietly nervous about

The feature getting the most love is also the one flagged in the safety writeup. Backchanneling, those “mhmm” sounds that make the model feel like it is actually listening, is exactly the behavior that makes a machine feel like a companion. GPT-Live-1 shows a regression on OpenAI's emotional reliance evaluation, from 0.88 to 0.82, which the company notes is not statistically significant while also committing to longer-term post-launch monitoring on that exact axis.^[2] Meanwhile early users are complaining the thing is too enthusiastic, and that the filler words get grating.^[10]

Both of those complaints are the same complaint. The model is optimized to feel present. Present is a hell of a drug, and the thing you optimize is the thing you get.

What I'd actually watch

• The API date and the per-minute price. Nothing else matters commercially until developers can build on it. The consumer launch is a demo of what the enterprise product will do. Watch what they charge, because that price sets the floor for every voice startup on earth.
• Published latency numbers. OpenAI shipped capability benchmarks and preference rates but no first-audio latency figure.^[9] For a product whose entire pitch is timing, that is a conspicuous omission. Moshi published 200 ms in practice.^[7] If GPT-Live is better, say so.
• Whether the backchannels get a volume knob. Trivial on its face. Actually the tell for whether OpenAI treats emotional stickiness as a bug or a retention feature.

The reason I keep coming back to the word full-duplex is that it is one of the rare cases where the marketing term and the engineering term are the same thing, and the engineering term is the interesting one. Voice assistants were not bad because the models were dumb. They were bad because they were built as request-response systems, and conversation is not a request-response system. Somebody finally fixed the topology instead of the model. That is usually where the real gains hide.