Tools / Analysis
GPT-Live-1.1 keeps a call open between questions
OpenAI added standby billing to its live voice API. A conversation can stay connected while the voice model pauses, leaving developers to decide when a pause is long enough to matter.

A voice assistant can finish giving a direction while the listener is still walking. It might be another minute before they need help. During that silence, there is little for a speech model to do, but closing the connection would force the listener to start again. OpenAI’s November 19 update gives developers another choice. For browser and app conversations using WebRTC, standby can pause the model during inactivity while the call stays open. Inference resumes when someone speaks.
What happens when you interrupt
OpenAI introduced standby alongside GPT-Live-1.1, its new full-duplex voice model. Full duplex means the model can keep listening while it speaks. In a conversation where someone interrupts to correct a name or switch languages, that lets the application receive the interruption on the same audio connection. OpenAI says version 1.1 improves pronunciation, accent consistency and transcription in multilingual conversations. Its release announcement describes the improvement. It does not give a per-language error rate.
There are two changes to test separately. The new model can use an existing Live session configuration, with instructions in ordinary language describing when to hand a task to a backend. An app might let the voice model handle an exchange about dates, then hand a request to check a booking to the system that actually has the records. Standby is a separate session feature available on both GPT-Live-1 and GPT-Live-1.1 over WebRTC. The developer can evaluate inactivity without migrating the voice model first.
Sources: OpenAI [1] · OpenAI [2] · @OpenAIDevs / X [3]
GPT-Live-1.1 release demonstration
Loading post from X…
The publisher’s example accompanies the voice API update.
The cost of silence
The distinction shows up on the bill. OpenAI lists 5 cents per active minute and 2 cents per standby minute, with backend reasoning and tools billed separately. A hypothetical ten-minute session with four active minutes and six in standby would use 32 cents of voice-session time. At the active rate throughout it would use 50 cents. A voice app that spends nearly the entire call talking will see much less difference. The important production measure is how many minutes the actual call spends in each state, and how the assistant responds when the person comes back.
Sources: OpenAI [1] · OpenAI [2]
SOURCES & METHOD
The record behind the article
Checked 21 November 2026. Technical claims are grounded in developer documentation; videos are credited to their creators. Company claims and third-party reviews are attributed in the text; conclusions are editorial analysis. No original comparative testing of these releases was conducted for this article.
- API changelog OpenAI
- GPT-Live-1.1 model and pricing OpenAI
- GPT-Live-1.1 release demonstration @OpenAIDevs / X


