OpenAI released GPT-Live-1 in the API on September 10, 2026, bringing its full-duplex voice model to developers. It can listen and speak simultaneously, and hand reasoning or actions to a separate backend. The announced voice-layer price is $0.05 per minute. This is the developer release following the earlier GPT-Live launch in ChatGPT. OpenAI’s launch announcement
The practical question is how well an application handles a changing request. Imagine someone asking to move an appointment, then correcting the date while the system is checking availability. A pleasant voice helps, but the application still needs to carry that correction through to the booking.
Conversation and execution have different responsibilities
GPT-Live handles conversation and decides when to delegate. The backend receives the work, uses tools, and returns results to be spoken. OpenAI documents two paths: Responses delegation for a supported Responses model, and client delegation for an application-controlled model, agent, or service. Realtime uses one model for speech, reasoning, and tool selection. GPT-Live getting started
That separation gives a team a useful design choice. An existing scheduling service could keep its business logic while GPT-Live becomes the spoken interface. A more involved research request could go to an agent. These are implementation examples, not integrations that arrive ready for a particular business.
OpenAI also lists native transcripts, response text, keyword biasing, and prompt control over tone and pace. Those features deserve separate evaluation: a voice can sound convincing while getting an account identifier wrong. Launch details
An interruption does not cancel a transaction
The most consequential line in the getting-started guide is that interrupting speech does not automatically cancel backend work. Applications own permissions, confirmations, private function execution, and durable task state. Application responsibilities
For the appointment example, “wait, make that Friday” should lead to an explicit decision about the pending booking. The application needs to know whether it is still searching, awaiting confirmation, or has already committed the change. Stopping audio alone cannot answer that question.
For an initial pilot, use a workflow whose outcome you can inspect. Record the requested change and compare it with the saved result. Include corrections spoken during tool execution, a repeated request after a connection failure, and a user who pauses without intending to finish. Measure successful outcomes alongside conversational quality.
Five cents per minute is the voice component
At the announced rate, 1,000 minutes would cost $50 for the voice layer. This is arithmetic using the launch price, not a total deployment estimate. OpenAI’s guide says duration is billed per second and backend model and tool use are separate. Telephony and application infrastructure should also have their own budget lines where applicable. Launch pricing · Usage accounting
The useful comparison for a pilot is cost per correctly completed request. A cheaper conversation that creates a booking error adds work somewhere else. Equally, paying for a more capable backend on every simple status lookup may be unnecessary. Compare a small, representative set of tasks before selecting a default.
Where to begin
OpenAI provides WebRTC for browser applications and WebSockets for server audio integrations, with telephony guidance available separately. Its browser quickstart requires a trusted server and keeps the project API key there. Connection options
Our recommendation is to test GPT-Live against the voice workflow you already understand, with the same backend permissions and success criteria. Give interruption handling a dedicated test rather than treating a clean demo as sufficient evidence. The new Agents API offers another approach to backend execution; deciding how that work is authorized remains part of the application design.