OpenAI announced API access to GPT-Live-1.1 on November 20; its technical changelog dates the change November 19. The model supports full-duplex conversation and backend delegation. WebRTC sessions can enter standby after configured silence: published rates are $0.05 per active minute and $0.02 in standby. Backend models and tools are charged separately.

Documentation describes Responses delegation to a supported model or connection to a custom backend. The private API key stays on a trusted server. This separates conversation from task execution without removing the application’s responsibility for permissions.

Conversation is not execution confirmation

With this architecture, an assistant can speak while a check remains underway. Our editorial conclusion is that the interface should distinguish an intention, an operation in progress and a verified completed result. Completion language belongs after confirmation from the system that actually performed the action.

For example, someone checking an order may add a detail during the conversation. If the backend is already working with earlier information, the application must decide whether to cancel, amend or complete that check. Otherwise a fluent conversation can conceal a mismatch between the person’s request and the task being executed.

Our proposed pilot

We would start with read-only access to test orders. Check interruptions halfway through a sentence, a corrected number and a long pause. Separately inspect what an operator sees when the backend is slow or returns an error. These are proposed scenarios, not tests of the model conducted by SIExpo.

Costs should be separated into conversation time, standby time and delegated work. A lower rate during quiet moments does not determine the final bill. A long-running task repeated after interruption may change the overall cost even when the voice session appears inexpensive.

Before allowing changes to real orders, we would require separate confirmation and a checkable operation identifier. After a dropped connection, the application needs to establish whether the write occurred before repeating it. A useful voice service must handle these awkward cases as well as produce natural pronunciation.

The operator should also be able to take over with the current task state intact. Hearing only the final spoken sentence may omit the uncertainty that prompted escalation. A concise record of the request, pending work and confirmed result gives human support a better starting point.

The API offers a more flexible structure for conversational applications. Practical readiness depends on state management, error handling and the actual cost of the complete workflow.