OpenAI has opened GPT-Live-1.1 to developers through its API, adding a standby mode that bills idle voice sessions at $0.02 a minute instead of $0.05. The full-duplex voice model listens while it speaks, and hands reasoning and tool calls to a backend model the developer chooses.
What standby changes
Standby works on WebRTC sessions. Developers set a speech-based wake policy and an idle timeout of five to 60 seconds, and the session pauses inference when nobody is talking. The connection, session ID and conversation history stay intact, and speech wakes the model again.
The saving is real for voice agents that spend much of a call waiting. Take a 30-minute support session with ten minutes of talk and 20 minutes of quiet: kept active it costs $1.50, but with standby it costs $0.90, since idle minutes bill 60 percent lower.
The model is live now through the v1/live/sessions endpoint, and developers can also reach it through the Realtime, Chat Completions and Responses endpoints. Day-one access includes the Vercel AI Gateway and OpenAI’s own Live playground, so teams can test it without wiring up a full stack.
What is new in 1.1
The 1.1 model also improves on GPT-Live-1 itself. OpenAI says it transcribes speech more accurately, holds a steadier accent and voice across languages, and completes tasks more reliably. It brings 13 new voices, and it has powered ChatGPT Plus and Pro voice since 15 October.
The model does not do the thinking itself. It handles the voice session while a backend model the developer picks does the reasoning and tool calls, and it keeps talking while that backend works. Switching from GPT-Live-1 is a model-name swap, and the API adds 38 voice IDs.
The limits to know
Standby has limits developers should plan around. It applies only to WebRTC sessions, not WebSocket connections, and the first reply after a pause can take longer. OpenAI has published no benchmark numbers for the transcription or accent gains, so those claims remain vendor claims for now.
Our Read
The price is the smaller part of the change. Voice agents spend most of a call listening, not talking, and until now every quiet second billed at the full rate. Standby turns that idle time into a configuration choice, and the saving compounds across thousands of calls.
The comparison that matters is with OpenAI’s own Realtime API, which bills audio by the token rather than the minute. Live is the simpler product: one voice model, one endpoint, per-minute billing. For teams whose agents mostly wait on humans, that simplicity now comes with a standby discount built in.
Why It Matters
Voice agents are moving from demos into real phone lines, and their economics decide whether they survive there. A model that charges for silence forces developers to disconnect between turns and rebuild context each time. Standby removes that tax, and it makes always-on voice agents cheaper to run at scale.
Background
The release caps a busy stretch for OpenAI. GPT-6.2 Astra and Sol launched on 17 November, and the company published a candid postmortem of the 21 November outage that knocked ChatGPT offline for three and a half hours. Voice is now the third front in the same week.
Source: OpenAI developer forum announcement and GPT-Live-1.1 rollout post


0 Comments