OpenAI releases GPT-Live-1 in the API for full-duplex voice

OpenAI made GPT-Live-1 generally available in the API on 10 September. The model supports full-duplex conversation: the voice interaction can continue while a backend model or agent handles reasoning and tools.

Developers can use an OpenAI backend through Responses delegation or connect their own backend through client delegation. OpenAI also documents native transcripts and response text, turn detection and keyword biasing.

The voice layer costs $0.05 per minute, billed per second. Backend model and tool usage are charged separately. That distinction matters when estimating the cost of a customer-support call: duration is one component, while reasoning, retrieval and business actions add their own charges.

Start with a bounded task such as checking an order status. Test a caller who interrupts, pauses, corrects a reference number or asks an unrelated question while a lookup is still running. Ensure the backend can cancel or supersede stale work, and require a clear confirmation before any action that changes customer data.

Keep the spoken interaction and the business action in separate logs, with a shared session reference. A pleasant conversation does not prove that the correct account was retrieved or that an update succeeded. Review transcript handling and retention before exposing a voice service to customers.

Measure cost per resolved call, successful tool execution and operator escalation alongside response time. Those measures make it easier to compare a voice-first experience with an existing support flow. These are our integration recommendations, not a report of a completed AZ Labs voice deployment.