
OpenAI has released GPT-Live-1.1 in its API, giving developers an updated voice model for conversations that continue while a backend agent works. The November 19 update improves pronunciation, accent consistency and transcription across languages, according to official OpenAI documentation.
Conversation and backend work run separately
The model can listen while it speaks and hand reasoning or tool-based tasks to a separate backend. Developers can select gpt-live-1.1 in an existing Live session configuration—an upgrade to the existing platform, not an entirely new agent system. The model documentation describes the multilingual update.
That split lets the voice layer keep a conversation going while a backend looks up information or runs a workflow. The backend model or agent is chosen independently of the voice model, as explained in the getting-started guide.
Standby keeps the connection, not active inference
The other practical addition is standby. Both GPT-Live-1 and GPT-Live-1.1 can keep a WebRTC connection open while pausing inference during inactivity, then resume when someone speaks. Developers can configure an idle timeout of 5–60 seconds. See the session-management documentation for configuration and lifecycle details.
Voice sessions cost US$0.05 per active minute or US$0.02 per standby minute. Backend model and tool usage are billed separately, according to the API pricing page. Standby makes connected pauses less expensive—not free. The lower rate applies to paused inference in a configured WebRTC session, not every moment when nobody is speaking.
Why it matters for builders
For builders, this separates the conversation from the work: users can keep talking while an agent checks information or runs a workflow. Better multilingual transcription is worth evaluating on names, numbers and mixed-language requests, not just fluent-sounding demos.
For SunMarc App Labs, a useful experiment would be a narrowly scoped voice interface for a utility workflow, evaluated on multilingual accuracy, task completion and total session cost. We have not tested this release; this is a proposed evaluation, not an existing SunMarc feature or a hands-on result.
The backend choice also affects the outcome. For context on OpenAI's separate reasoning and coding models, see our GPT-6.2 Astra and Sol coverage.