OpenAI introduced GPT-Live-1 in its developer API on September 10, extending a voice model first offered in ChatGPT to custom applications and business workflows. It can listen and speak simultaneously, allowing callers to interrupt, add information or keep talking while a separate backend works on a task.
The distinction is more than voice quality. GPT-Live-1 handles the conversation; another model, agent or service handles deeper reasoning and actions. OpenAI’s documentation uses an order-status example: a caller can add a detail while the backend checks the order, and the voice model can explain the result when it arrives.
Developers can steer the voice layer’s tone, pace and style through a system prompt. OpenAI says the model also handles silence and background noise and provides transcripts and text output.
Two ways to connect the work
With Responses delegation, OpenAI runs a supported backend model and passes conversation context and results between it and the voice model. The application still executes its own function tools. Client delegation instead connects an existing model, agent harness or service and gives the application control over backend context and execution.
The delegation mode is chosen when creating the session; changing it requires a new session. Detailed business rules and tool workflows belong with the backend, while the voice prompt governs conversation and when to request help.
Browser integrations can use WebRTC, with a trusted server holding the API key. WebSockets support server-side audio integrations, and the documentation also covers telephony and SIP. A backend can also monitor or control a session through an additional sideband WebSocket while audio continues over the primary connection. OpenAI lists integrations for existing voice stacks including LiveKit, Twilio, Telnyx and Daily/Pipecat.
Billing and control remain separate
At launch, the voice stream costs $0.05 per minute, billed per second. Backend model and tool usage is charged separately, so the voice rate is not a complete price for an agent-handled call.
The application must check permissions, obtain required confirmations, execute functions against its systems and save durable task progress. Interrupting the spoken response does not automatically cancel backend work. Developers must decide whether that work should finish or stop, an important distinction when a conversation can continue while actions are still underway.