OpenAI has made GPT-Live-1 generally available in its API, and its most consequential feature is not a new voice. It is the refusal to stop listening while the rest of the system thinks.
Traditional voice agents tend to move a call through a relay race: speech-to-text, a language model, then text-to-speech. A caller pauses, interrupts, corrects themselves or speaks to somebody else, and the handoffs become obvious. GPT-Live-1 instead handles incoming and outgoing audio together, while delegating deeper reasoning and tool calls to a backend model. OpenAI says that arrangement improves turn-taking and lets an agent continue the conversation while work happens elsewhere.
That is a meaningful architectural change for companies building support lines, tutoring products and voice-controlled tools. It is not, by itself, a production verdict. The voice layer costs $0.05 per minute, billed by the second, but the model behind it, tool calls, telephony or real-time communications and the operations around a real call remain separate costs. The practical question is no longer whether a demo sounds natural. It is whether a resolved call becomes cheaper, safer and more reliable.
A voice front end, not a complete agent
OpenAI positions GPT-Live-1 as a full-duplex front end. The model can listen and speak continuously, respond to interruptions and acknowledgements, handle silence and background noise, and produce ASR transcripts and response text. Developers can use OpenAI Responses delegation with an OpenAI backend or client delegation with their own. The company also says developers can pair the voice model with a third-party model.
This unbundling is useful. A scheduling call may need a fast, inexpensive backend; an insurance exception may need more capable reasoning; a regulated action may need a deterministic rules check and a human. It also means that swapping a chained stack for one voice model does not remove orchestration. It moves the most consequential orchestration—the authorization, retry, state and escalation logic—behind the audio session.
The $0.05 rate is a floor, not a call price
The listed voice charge is refreshingly legible: five cents per minute. It is also only the front door. A product team needs to add backend inference, tool usage, carrier or WebRTC charges, call recording and storage where applicable, observability, and the human minutes spent on escalations. A low apparent voice rate can be overwhelmed by a slow or overly capable backend, repeated tool calls, or a design that asks people to repeat themselves.
| Layer | What GPT-Live-1 changes | What a buyer still has to measure |
|---|---|---|
| Conversation | A single full-duplex voice model can reduce handoffs between STT, LLM and TTS. | False end-of-turns, interruption recovery, transcription accuracy by accent and noise, and response latency. |
| Reasoning and tools | The voice session can delegate work while keeping the interaction moving. | Tool correctness, retry behavior, backend latency, context transfer and per-resolution spend. |
| Safety and operations | OpenAI describes live input/output checks and interruption or termination controls. | Authorization boundaries, audit logs, handoffs, retention, incident review and compliance obligations. |
Source: OpenAI’s GPT-Live-1 announcement and release notes, September 10, 2026. The rightmost column is AI News Fab analysis, not an OpenAI benchmark.
Benchmarks point in the right direction, with limits
OpenAI reports that GPT-Live-1 improves its Full Duplex Bench performance by 30 percentage points over GPT-Realtime-2.1. It also says GPT-Live-1 paired with GPT-6 Astra at medium reasoning ranked first on Tau3, a spoken banking-support benchmark using knowledge retrieval and account tools. Those claims fit the architecture: fewer turn boundaries should help with pauses, barge-ins and self-corrections.
But a benchmark that rewards a tool sequence or a turn-taking response does not settle whether a specific support line will work. Production calls add line noise, code-switching, interruptions from another person, long holds, inconsistent data, failed API requests and moments when a confident agent must not act. A team moving from a cascaded build should run its own controlled comparison rather than assuming OpenAI’s result transfers.
Smoother speech does not grant authority
The system card describes voice-native evaluations, live checks of inputs and generated outputs, and controls that can steer, interrupt or end unsafe conversations. It does not decide whether an agent may cancel a reservation, disclose account information, change a prescription or trigger a refund. Builders still need explicit approval thresholds, scoped tools, idempotent actions, confirmation design, audit trails and a fast handoff to a person.
Who should evaluate it now
GPT-Live-1 is compelling for teams whose current pain is visibly conversational: tutoring that interrupts learners, phone systems that lose the thread when a caller speaks over them, field-service tools used with hands occupied, or support flows where natural back-and-forth improves information capture. It is less urgent for a product with short, tightly scripted calls that already have good speech recognition and deterministic backend logic.
Start with a small, representative call set. Segment it by noise, accent, caller intent and tool complexity. Record interruption recovery, word-error rate, first-response time, tool-call correctness, abandonment, escalation success and total cost per resolved call. Only then compare GPT-Live-1 with the existing chained stack or another real-time API.