The Full-Duplex Race — GPT-Live Opens a New Front in the Voice-Model War

On July 8, 2026, OpenAI released GPT-Live, a pair of voice models — GPT-Live-1 and a smaller, faster GPT-Live-1 mini — built on a full-duplex architecture that listens and speaks at the same time. Rolled out inside ChatGPT to a global user base, the models can hold a continuous conversation, back-channel with small sounds like "mhmm" while the user is still talking, pause when the user is thinking, and interrupt or be interrupted without breaking. For the moment-to-moment exchange, GPT-Live-1 runs on GPT-5.5 in the background and delegates heavier reasoning and web search to the latest frontier model, folding the result back in when it is ready. OpenAI has said the developer API will follow.

The launch is significant less for being first — it is not — than for where it landed and how completely it reframes the contest. Full-duplex speech-to-speech had been demonstrated by Kyutai with Moshi, which reaches end-to-end latency around 200 milliseconds by modeling user and system audio as parallel streams, and the major cloud and model vendors have been shipping real-time voice for the better part of a year. What GPT-Live does is drop a frontier-grade, natively full-duplex model into the single largest consumer AI distribution channel that exists, which turns a research direction into a mass-market default overnight.

A field that filled up fast

The competitive picture, assembled from vendor announcements and independent benchmarking through the first half of 2026, is already dense. Microsoft has released GPT-Realtime 1.5 and a new Azure-Realtime model in public preview through Azure AI Foundry, pitched at more natural voice output across languages and accents. xAI has a Grok Voice Agent that, in early-2026 measurements, posted the fastest time-to-first-token of the major systems. Amazon has Nova 2 Sonic. Google has Gemini 3.1 Flash Live. NVIDIA released PersonaPlex out of its ADLR group, adding voice and role control to full-duplex conversational speech. The approaches split into two architectural camps — cascaded pipelines that stitch speech-to-text, a language model, and text-to-speech together, and native speech-to-speech models that collapse all three into one — and 2026 is the year the second camp moved from research to product.

The through-line across all of them is that the interesting number changed. Time-to-first-token, the metric the field has been ranking itself on, measures how fast a reply begins — which is a cascade concern that still assumes the conversation happens in turns. As of early 2026, measured end-to-end time-to-first-token ranged from well under a second to nearly three seconds across platforms, a spread wide enough to separate systems that can hold a real conversation from systems that cannot.

Even the fastest production systems sit three to four times above the roughly 200-millisecond gap that separates one human speaker from the next, which is why a natively full-duplex model — one that can begin a back-channel almost immediately while the substantive answer is still forming — is a categorically different proposition from a cascade that has merely optimized its first token. The full analysis of why the turn-based pipeline is being abandoned, and what that does to the voice stack and its economics, is laid out in the companion feature on why full-duplex voice ends the era of taking turns.

The two-brain pattern goes mainstream

The design choice most worth watching is the split between a fast reflex model and a slower reasoning model. GPT-Live keeps the conversation alive on a quick model and escalates to a frontier model only when a turn genuinely needs it. This is not merely an engineering nicety; it is the economic model that makes always-on voice affordable. A voice session is a long stream of cheap, small decisions punctuated by occasional expensive ones, and paying frontier prices for every millisecond of listening would make the product uneconomic. Metering the expensive capability instead of the whole session is the same consumption-billing logic reshaping the rest of enterprise AI.

Why the scoreboard is being rebuilt

The uncomfortable subtext of the launch is that the metrics the industry has been competing on do not capture full-duplex quality. Word error rate measures transcription, a cascade artifact. Time-to-first-token measures reply speed, which still assumes a turn. Neither measures whether a model back-channels at the right moment, yields gracefully when interrupted, holds a pause instead of filling it, or handles a tool call correctly when the user corrects themselves mid-sentence. The research community has responded with a new class of benchmarks built specifically around full-duplex behavior, real-world disfluency, and tool use under interruption — a sign that even the people who measure this for a living agree the old scoreboard is measuring the wrong game.

The enterprise read

For companies deploying voice, the launch is a planning signal, not a gadget. The first implication is that the conversational-cascade architecture is becoming legacy for real-time use; teams with roadmaps full of endpointer tuning and barge-in middleware are polishing components that native full-duplex removes. The second is that acceptance tests written for cascades will pass a full-duplex model that is subtly unpleasant to talk to, because those tests measure transcription and latency rather than floor-sharing — so evaluations have to be rebuilt before migration, not after. The third is that always-listening, by definition, changes the privacy posture of any voice deployment, and the norms for continuous audio processing lag the capability badly.

There is also a hardware dimension that decides who can serve this at scale. A full-duplex model is a hard real-time system that demands low-variance, low-latency inference for the length of every conversation, which is exactly the workload the current wave of purpose-built inference silicon is being tuned for. That connection was the subject of the earlier reporting on the inference-silicon war and the shift to purpose-built chips, and it is why a company with a good voice model but no low-latency serving substrate cannot ship real full-duplex at scale.

How fast this consolidates is the open question, and it is one worth committing to in writing. The dated, falsifiable marker for when full-duplex becomes the default for real-time voice APIs lays out the specific conditions to grade against. The short version: the models are converging faster than the benchmarks, the norms, or the tooling around them, and the winner will be decided less by whose latency is lowest in a given week than by whose voice is already in the room when the user starts to speak.