Every voice assistant ever shipped has worked like a walkie-talkie. You speak, you stop, it detects the silence, it answers. On Thursday 10 September, OpenAI put a model into the API that breaks that contract. GPT-Live-1 listens and speaks at the same time — the way people on a phone call do — and it costs $0.05 per minute for the voice layer, billed by the second, with whatever backend model you attach charged separately at its normal rates.
The distinction matters more than the marketing term full-duplex suggests. Conventional voice agents, including OpenAI's own Realtime API line, are half-duplex: they chain speech-to-text, a reasoning model and text-to-speech, and they rely on voice-activity detection to guess when your turn has ended. Every handoff adds latency, and VAD has no idea whether 900 milliseconds of silence is a finished sentence or a person thinking. GPT-Live-1 collapses the stack into a single model that reasons over incoming and outgoing audio together, which OpenAI says avoids the latency and brittle handoffs of chained STT–LLM–TTS architectures. It still exposes turn detection for developers who want explicit boundaries. It simply no longer needs them.
The numbers OpenAI published are large enough to invite scepticism and specific enough to check. On Tau3 Voice Intelligence, which scores spoken customer-service tasks across airline, retail and telecom domains, GPT-Live-1 posts 86.2% Pass@1 against 45.7% for GPT-Realtime-2.1 and 42.4% for GPT-Realtime-2, paired with GPT-6 Astra at medium reasoning effort. On Full Duplex Bench, OpenAI reports a 30-point gain over GPT-Realtime-2.1, with turn-taking latency falling from 1.41 seconds to 0.798 seconds and interactivity climbing from 45.4% to 80.10%. That is not an incremental release; it is a category reset in one generation.
Launch customers are unusually specific about what changed. Language-learning app Speak said the model stopped talking over its students: in early evaluations, co-founder and CTO Andrew Hsu reported, it “cut interruptions during thinking pauses by almost 80% compared with previous turn-based systems.” Yelp, which routes reservations and food orders through its Yelp Host product, framed the gain commercially. “With GPT-Live-1, every call is more conversational and responsive,” said chief product officer Akhil Kuduvalli Ramesh, who credited the model with capturing revenue restaurants would otherwise have missed while staff were tied up on the phone. A health-care developer quoted in OpenAI's launch post put the engineering saving bluntly: moving off a cascaded build simplified the code base by 80% and removed 23,000 lines of code.
The architecture is the other half of the story. GPT-Live-1 does not do the thinking. It holds the conversation while delegating reasoning and tool calls to a backend the developer picks — GPT-6 Astra for complex issues, a cheaper model like Luna or Terra for scheduling and order lookups, Codex for engineering work, or a third-party model entirely. It ships with twelve voices spanning accents and languages, plus native ASR transcripts, keyword biasing, alphanumeric handling for reading back account numbers, and telephony support for real phone lines.
Why barge-in was the last uncanny valley
Voice AI solved intelligibility years ago. What it never solved was timing — and timing, in speech, is meaning. A pause is input. A murmured acknowledgement is a backchannel, not a new turn. Cutting someone off mid-sentence is the most reliable way to make a machine feel like a machine, and every half-duplex system does it constantly because silence is the only signal it has.
Independent measurement suggests OpenAI bought that naturalness with a trade, not a free lunch. Agora's media lab, which tested GPT-Live-1 in the ChatGPT app in July using an artificial mouth and waveform-level timing, found the model is actually 498 ms slower to go quiet when you deliberately interrupt it — it spends that half-second deciding whether you meant it. The payoff showed up elsewhere. GPT-Live-1 rejected all thirty background-speech probes that stopped the previous turn-based mode twenty times, and its response-latency variance collapsed roughly fivefold, from a standard deviation of 489 ms to 104 ms. That is the sharper insight: the median barely moved, but the delay became predictable, and predictability is what brains read as natural. The lab also logged a new failure mode — in a noisy room where nobody addressed it, the model spontaneously answered background voices in four of thirty windows. Agora sells real-time transport, so discount the framing accordingly; the methodology remains the most rigorous public read available.
Then there is the price. At $0.05 a minute, the voice layer costs $3.00 an hour — before the backend model, the agent harness and telephony. That undercuts or matches most of the field on headline rate: Deepgram lands near $0.056/min all-in, Hume runs $0.04 to $0.07, Retell starts around $0.07, and ElevenLabs and Grok Voice sit near $0.08. But headline per-minute pricing here is close to fiction. Practitioners report real all-in costs running two to three times the advertised figure once reasoning tokens are counted, with OpenAI-based stacks landing around $0.15 to $0.20 per minute. Even so, that is a fraction of a loaded human agent hour — and it now arrives without the script-reading cadence that gave the game away.
The competitive squeeze is obvious. Google's Gemini Live went generally available at I/O 2026, Amazon has Nova Sonic, and a whole middle layer of orchestration vendors — Vapi, Retell, LiveKit, Pipecat — exists to paper over seams GPT-Live-1 just removed.
Three things to watch. Whether third-party evaluators reproduce the Tau3 and Full Duplex Bench gains outside OpenAI's own harnesses, since the backend pairing does much of the work in those scores. Whether the speaker-attribution problem Agora surfaced — an agent joining conversations it was not invited to — becomes a support liability once these systems answer real phone lines at volume. And whether the orchestration layer consolidates or simply gets absorbed. The stop-start rhythm was always the tell. If it is genuinely gone, the next argument is not about whether callers can detect that they are talking to an AI. It is about whether anyone is required to tell them.
“Cut interruptions during thinking pauses by almost 80% compared with previous turn-based systems.”— Andrew Hsu, Co-founder and CTO, Speak