OpenAI has made GPT-Live-1.1, the voice model behind voice chats in ChatGPT Plus and Pro since 15 October, available in its API as gpt-live-1.1. It announced the release on its developer forum on 20 November, together with a standby option for both GPT-Live-1 and GPT-Live-1.1. Standby pauses the model while nobody talks and bills those minutes at $0.02 instead of $0.05, 60% less.
GPT-Live listens while it speaks, so users can interrupt it, and it can hand harder questions or tool calls to a backend model you choose. OpenAI says version 1.1 transcribes speech more accurately and keeps a more consistent accent across languages. It also claims better handling of letters and numbers and more reliable task completion, but publishes no figures for any of this.
Standby pauses the model after 5 to 60 seconds without speech
You switch standby on per session, as OpenAI’s guide to configuring standby explains: set the wake policy to speech and pick an idle timeout between 5 and 60 seconds. When no one has spoken for that long, the model pauses, but the WebRTC connection and the session ID stay the same. The microphone stays live, and when the user speaks again the model wakes up with the conversation history intact.
Your app gets an event at each pause and resume, so it can show whether the assistant is listening. The first reply after a short remark can take a little longer, OpenAI notes. Leave the timeout out and the model stays active, at the full rate, until the session ends.
Active minutes cost $0.05 and standby minutes $0.02, with the backend model billed on top
GPT-Live-1.1 costs $0.05 per minute while the model is active and $0.02 per minute in standby, according to OpenAI’s model page. The backend model that handles delegated questions, and any tools it calls, are billed separately.
The difference adds up when a session stays open for hours. Take a help kiosk that is connected for an eight-hour shift and hears speech for two of those hours. Fully active, the voice layer costs $24 a day. With standby, our calculation comes to about $13, plus a little for the seconds of silence before each pause.
GPT-Live-1.1 also has its own cap on simultaneous sessions, separate from GPT-Live-1. OpenAI does not publish its size per usage tier, so check your organisation’s limits page before a large pilot.
Standby covers browser and app connections, and the voice list leaves out Dutch
OpenAI describes standby only for WebRTC, the connection type browsers and mobile apps use for live audio. GPT-Live also runs over WebSocket, but the documentation offers no standby there and does not mention phone lines. If your voice agent answers calls through a telephony provider, confirm that the lower rate applies before you count on it.
Language is the second thing to check. OpenAI’s voice guide lists voices for English, Brazilian Portuguese, German, French, Hindi, Japanese and Korean. Dutch is not named. The model is described as multilingual, so it will speak Dutch. Nothing OpenAI published shows how well it handles Dutch names, postcodes or account numbers.
European data residency lists GPT-Live-1, not yet version 1.1
OpenAI’s data residency page lists GPT-Live sessions for Europe, with storage and processing in the EEA and Switzerland, but names only gpt-live-1 as the model. If voice data must stay in Europe, GPT-Live-1 with European data residency is the documented route for now. Standby is available for that model too. OpenAI adds 10% to the price of data residency endpoints for models released since 5 March 2026, which includes GPT-Live-1.
For privacy teams: GPT-Live sessions are eligible for zero data retention, and session storage is off by default. The backend model that GPT-Live delegates to has its own data controls, so its location needs a separate check.
Assistants that stay open in kiosks, apps and field work gain the most
Standby matters for voice assistants that stay connected while people work or wait. Think of a help point at a hospital entrance, a support widget in an app, or a technician’s hands-free assistant. Those sessions are mostly silence, and silence now costs 40% of the active rate. For short customer calls that end when the caller hangs up, little changes.
For the Dutch organisations DataNorth works with, language and data location still decide more than the price. A voice agent that mishears a postcode creates work instead of saving it. Mistral’s Voxtral 3, launched in October, lists Dutch voices and runs on an EU-hosted API from $0.04 a minute, with Mistral Small 4 included for reasoning. For Dutch-language projects it is the obvious comparison, although the two prices do not cover the same parts.
Building a voice agent for a browser or app? Switch on standby in a test environment now and measure how often it pauses and how fast it wakes. Run test calls in Dutch with real names, postcodes and order numbers before GPT-Live-1.1 talks to customers. If your data must stay in Europe, keep GPT-Live-1 for now and wait until OpenAI adds version 1.1 to its European data residency list.
For more information, visit the official announcement of GPT-Live-1.1 on the OpenAI Developer Community forum.