GPT-Live API: voice conversations with a separate backend

GPT-Live 1 became generally available in OpenAI's API on September 10, 2026. Its conversation layer can listen and speak while a separate backend handles reasoning and tools. The listed voice-session charge is $0.05 per minute, billed per second, with backend model and tool charges separate. API announcement.

The architectural change matters when a conversation needs to continue while work is still running. It also means an interruption in speech and cancellation of a business action are different events.

A correction must reach the work, not just the conversation

Imagine a creator asking a voice assistant to prepare a campaign report for August, then correcting the month to September while the data is loading. This is an illustrative scenario. A useful system acknowledges the correction and produces the September report. A pleasant voice that announces the original August result has failed the task.

The application needs to track which request the backend is completing and decide whether to cancel, replace or disregard outdated work. OpenAI's guide explicitly leaves permission checks, task state and interruption decisions with the application; backend work can continue after the caller interrupts. GPT-Live architecture.

Test that case deliberately before promising an unattended assistant. Delay the report lookup, correct the month, and inspect the final document. The acceptance criterion is one report for the requested period, with no outdated result presented as current. This is a proposed test, not an observed product result.

Two ways to connect the backend

Responses delegation uses an OpenAI model for the backend, while your application still executes its own function tools. Client delegation connects an existing model, agent or service that your application controls. Keep conversation style in the voice instructions and detailed business rules with the backend. Delegation overview.

For the campaign example, that division helps locate mistakes. Hearing the wrong month is a conversation problem; querying the correct month but calculating the wrong total is a backend problem. The final report provides evidence about what happened beyond the spoken acknowledgement.

Budget for the complete session

Twenty minutes at the published voice rate costs $1 for that component. Ten such sessions cost $10 before backend reasoning, tools and infrastructure. Those are arithmetic examples, not deployed-agent quotes. Repeated or abandoned backend work can add costs even when the user receives only one report.

Keep a session-level record of voice duration, backend usage and completed outcomes. Compare the total with the existing workflow before deciding whether the conversational experience justifies the additional integration work.

GPT-Live does not make every Realtime application obsolete. Choose based on the conversation and backend behavior you need, then test corrections and slow operations against that design. Our voice comparison covers the alternatives. This report concerns the developer API, not universal access or pricing for consumer voice features.