OpenAI launches GPT-Live-1, a full-duplex voice model that delegates reasoning to another backend
On September 10, OpenAI brought GPT-Live-1 to its API: a full-duplex voice model that holds the conversation and hands heavy questions to a reasoning backend. For teams building voice agents, the win is measurable — 80% less code for one customer — but it is paid for in platform dependency.
September 10, 2026. OpenAI launched GPT-Live-1 in its API, bringing the native voice architecture behind ChatGPT voice mode to outside developers for the first time. Tony Stoyanov, CTO of EliseAI, sums up the payoff in one number: 23,000 lines of code deleted, about 80% of his codebase. The central mechanism fits in a sentence: one model holds the conversation, and the heavy reasoning is delegated elsewhere.
The classic voice stack is a fragile chain
A voice agent is almost never a single model. It is a chain of systems handing a conversation back and forth: speech is transcribed to text, a model decides on a response, then that text is synthesized back into speech. Every link adds latency, and the budget is unforgiving — human conversation tolerates delays badly beyond a few hundred milliseconds. The result sounds robotic fast, not only in the voice itself but in the mechanics of turn-taking.
GPT-Live-1 aims to collapse that stack. It acts as the conversational frontline: natively full-duplex, it keeps up with a conversation in real time, including when someone cuts in mid-sentence, without developers coordinating separate systems. The difference from the half-duplex of earlier generations — where each side talks in turn, walkie-talkie style — is exactly what changes the perception: an interlocutor who can be interrupted is a credible one.
Delegating the reasoning
The voice model does not have to do everything itself. When a request needs more time or processing, GPT-Live-1 hands it off to another model in the background: GPT-6 Astra, a smaller model like Luna, or a model from another provider entirely.
That is the answer to the silence problem that plagues voice agents: ask a hard question and sit in silence while the model thinks. GPT-Live-1 keeps the conversation going instead — filling pauses, acknowledging the speaker — then works the answer in once the backend is done. Reasoning becomes invisible to the caller, who perceives only a fluid conversation.
On the numbers, OpenAI claims 30 percentage points better than GPT-Realtime-2.1 on Full Duplex Bench, and first place on the τ³-benchmark paired with GPT-6 Astra at medium reasoning.
What delegation produces, concretely
Delegation runs through an event-driven interface: the voice session generates a delegation_id, sends context to the backend doing the heavy work, and gets the result back through the session.commentary.append event. The voice model folds that result into the ongoing conversation rather than reading a block of text aloud. Developers keep visibility into what the model hears and says, and control when it takes a turn — they just no longer have to assemble the whole conversation out of separate systems.
Early customers put numbers on the win.
- EliseAI, a healthcare platform, deleted 23,000 lines of code and shrank its codebase by 80%, time reinvested in the patient experience — scheduling and navigating care.
- Speak, a language-learning app, measured that GPT-Live-1 interrupts learners 80% less often when they pause to think. When learning a language, those extra seconds are the difference between getting the answer out and being cut off.
- Yelp is already using it in Yelp Host and Hatch: CTO Alex Levy reports more calls handled successfully by AI, and callers speaking in fuller, more natural sentences.
The price of the voice layer
GPT-Live-1 costs $0.05 per minute, about $3 an hour — on top of whatever reasoning backend it calls. Every time the agent reaches for GPT-6 Astra, the bill climbs with it. The more often an agent leans on a reasoning model, the faster the total cost grows.
OpenAI has been cutting API prices as competition from Anthropic, Google, and Chinese labs heats up, but frontier reasoning is still not free. The point of the decoupling is selectivity: a simple appointment booking can go to Luna, a question that actually needs multi-step reasoning or tool calls can go to Astra. GPT-Live-1 gives teams a place to apply that cost logic to voice, call by call.
What full-duplex means in practice
Full-duplex is not a spec detail. In a half-duplex system, the model waits for the user to finish before acting: it must detect the end of turn, trigger transcription, infer, then synthesize. Every step adds latency, and the system fails on interruption — barge-in, when the user cuts in. A full-duplex model listens and speaks at the same time: it can start answering while the user is still talking, and be interrupted cleanly. That is what separates a conversation from an alternation of monologues.
The gain shows up in cases where interruption is expensive — a call center, a medical appointment, a language lesson. Speak puts a number on it: 80% fewer interruptions when a learner pauses to think.
A launch demo showed the difference in a noisy setting: a restaurant reservation kept moving while background noise played and two people talked over each other. That is the bar voice agents now have to clear — not transcribing correctly in a quiet room, but holding a turn under real-world interference.
Voice becomes an arena
The cost is quick to compute. A ten-minute support call is $0.50 of raw voice layer; if the agent delegates half the time to GPT-6 Astra at medium reasoning, the bill climbs accordingly with the backend rate. The structure rewards teams that reserve reasoning for questions that deserve it, and punishes those who put a frontier model on every turn.
GPT-Live-1 does not arrive in a vacuum. Google, Anthropic, and Chinese labs are pushing their own real-time voice models, and OpenAI has been cutting prices under that pressure. Voice is becoming the next mass interface — not just for assistants, but for booking, support, and healthcare. In that race, the “one model talks, one model thinks” decoupling is the architecture that wins, because it optimizes cost and latency separately. GPT-Live-1 is the first major implementation of that decoupling exposed as an API.
The control tradeoff
With the older cascaded approach, every link in the stack could come from a different provider and be swapped out at will. GPT-Live-1 takes over more of the conversation — which also means handing more of the stack to OpenAI.
The bet is explicit: developers will give up some control if voice agents can finally keep up with the people talking to them. It is a reasonable bet, but it is one you sign knowingly.
Verdict
GPT-Live-1 is the first serious attempt to turn agent voice into a single API product rather than a hand-built assembly.
If you are building a latency-sensitive voice agent, the jump is quantifiable and justified: 30 points on Full Duplex Bench, 80% less code at EliseAI, and interruption nearly eliminated at Speak. Test it on a use case where interruption is expensive — booking, healthcare, learning. If you must stay multi-provider, measure the exit cost before signing: GPT-Live-1 locks more of the stack into OpenAI, and the per-minute price does not include the reasoning backend. The latency win is real; so is the dependency.