OpenAI made GPT-Live-1.1 available through its Live API in a November 19 changelog update, bringing its newer conversational voice model to developers’ applications. The release also adds standby for GPT-Live-1 and GPT-Live-1.1 WebRTC sessions, reducing voice charges while inference is paused without closing the call.
GPT-Live-1.1 handles speech while a backend does the work
Developers can select gpt-live-1.1 in an existing Live session configuration. The full-duplex model can listen while speaking and hand reasoning or tool work to a backend, keeping the spoken conversation separate from the system executing the task.
OpenAI reports improved pronunciation, transcription and accent consistency across languages compared with GPT-Live-1. Those are the company’s capability claims; the release materials inspected for this report do not establish how much a particular application’s task-completion rate will improve.
For builders of voice agents, the practical change is a new model choice within the existing session interface. The backend’s model, tools and execution logic still need their own configuration.
Standby preserves the call during quiet periods
Active voice sessions cost $0.05 per minute; standby costs $0.02 per minute, with voice usage billed per second. The lower rate applies to WebRTC sessions while inference is paused, rather than automatically applying to every silent moment.
Developers can configure an idle timeout from five to 60 seconds. After that interval without an input or output transcript, the session returns to standby. Speech wakes it again, with the connection and session identity retained and conversation history restored.
The microphone track must stay active for speech to wake the session. OpenAI cautions that the first response to a short utterance after standby may take longer. Omitting the idle timeout leaves inference active after the first wake until the session ends.
Backend charges and initialization affect the total
Model and tool usage for delegated work is billed separately from voice duration. Active time also includes silence before the timeout and waits for backend work while inference continues. Muting the microphone alone leaves the session active.
OpenAI’s cost guide says default WebRTC startup bills 15 seconds during initialization, credited against duration charges once the session runs. That is included in a running session’s duration, rather than an extra 15 seconds to add again.
For longer gaps, the documentation recommends considering closing the voice session and restoring context in a new one. The useful cost comparison therefore includes voice duration, backend work and reconnection behavior, alongside the delay a user experiences when resuming.