OpenAI Ships GPT-Live-1.1 Voice API With Cheaper Standby and Sharper Speech

OpenAI has brought GPT-Live-1.1 to its API, and the headline feature is not a new voice — it is a cheaper way to stay quiet. The model, which has powered ChatGPT Voice since October 15, now comes with a standby mode that drops idle time to $0.02 a minute, against $0.05 a minute while it is actively talking. For anyone running a voice agent all day, that 60% cut on silence is the whole story.
GPT-Live-1.1 arrived in the API on November 19-20, alongside the same standby mode for GPT-Live-1. It is a full-duplex voice model, which means it can listen while it speaks — a caller can interrupt mid-sentence instead of waiting for a turn to end. It also transcribes speech more accurately and holds a steadier accent across languages than its predecessor, according to OpenAI’s own model page.
How standby works
Standby is a billing state, not a disconnect. When a WebRTC session goes quiet, OpenAI pauses inference but keeps the connection, the session ID and the conversation history. Speech wakes it again. You configure it when you create the session, with two fields: wake_policy.type set to “speech”, which tells the session to wait for voice activity before starting inference, and idle_timeout_s — a whole number from 5 to 60 — which sends an active session back to standby after that many seconds pass with no transcript on either side.
Your app can watch for session.paused and session.resumed events on the data channel, so the interface can show whether the model is listening or resting. One detail to plan around: the microphone track has to stay open during standby, since the idle timer measures silence rather than a muted mic, and the first reply after a wake can take longer than a normal turn because inference is cold-starting again.
Standby applies to WebRTC sessions only. Phone calls and WebSocket connections still pay the full active rate, and the backend model and tool usage that GPT-Live-1.1 calls through delegation are billed separately and do not change with standby.
What it actually saves
Take a 30-minute support session with ten minutes of real conversation and twenty minutes of quiet — the customer looking something up, a form being filled in, a human agent doing work on their side. Kept active the whole time, it costs 30 x $0.05 = $1.50. With the quiet time in standby, it costs 10 x $0.05 + 20 x $0.02 = $0.90. That is a 40% saving on the session, and the gap grows with the share of silence: an eight-hour support line with mostly idle time saves far more.
Both rates are billed per second, so you only pay for the silence you actually have. The model also carries a 128,000-token context window, and it is live through the Live, Realtime, Chat Completions and Responses endpoints, so nothing about how you connect changes.
Who should care
The winners are the teams whose voice agents spend most of their time waiting: support lines, receptionist agents, sales qualification calls, and any assistant that sits on an open line while a human does something else. If your calls are mostly talking, standby changes little. If your calls are mostly waiting, it changes the bill by a lot.
Voice has been a busy track this month. We covered Amazon Polly’s voice cloning yesterday, and OpenAI has been pushing ChatGPT deeper into work tools — including businesses inside ChatGPT and computer-use agents with plugins. GPT-Live-1.1 is the developer-side piece of the same push: a voice model cheap enough to leave running, with the silence priced like silence.
If you are already building on GPT-Live, the change is a small one — two session fields, and a decision about your idle timeout. Set it short for support lines with long pauses, longer for conversations where people think out loud. And if you have been holding off on a always-on voice agent because the per-minute math never worked, run the numbers again with standby. The silence just got 60% cheaper.
