OpenAI's GPT-Live-1.1 Makes Voice Agents Cheaper to Leave Running
OpenAI put GPT-Live-1.1 in its API on November 19, and the headline is not the better voice. It is standby mode, which cuts the cost of an idle voice session by 60 percent.

OpenAI put the GPT-Live-1.1 voice model in its API on November 19, and the headline is not the better voice. It is the price of silence. A new standby mode bills idle WebRTC sessions at $0.02 per minute instead of $0.05, a 60 percent cut that matters a lot more than it sounds like, because voice agents spend most of their lives waiting.
GPT-Live-1.1 is OpenAI's full-duplex voice model, which means it listens while it speaks instead of taking turns like a walkie-talkie. It handles the speech itself and hands the reasoning and tool calls to a backend model you choose. You reach it through one endpoint, v1/live/sessions, over WebRTC, WebSocket, or SIP. The model ID is gpt-live-1.1, and switching an existing Live session to it takes a new session and one changed string.
GPT-Live-1.1 Voice Model: What Actually Improved
OpenAI lists four improvements in the changelog, and they are all about the voice, not the brain. GPT-Live-1.1 transcribes speech more accurately, holds a more consistent accent and voice across languages, and handles letters and numbers better in multilingual conversations. It also ships 38 voice IDs, including seven new ones with names like calandria, hornero, and luciernaga that lean toward Spanish-language use.
The model is not new to everyone. It has powered ChatGPT Plus and Pro voice since October 15, so the API release is really OpenAI letting developers use the voice its own paying users have been talking to for a month. Live sessions hold a 128,000-token default context window, and the model supports streaming and function calling, but not structured outputs or fine-tuning.
That split is the whole design. GPT-Live is a voice frontend, not a reasoning engine. The older GPT-Realtime family does more of the thinking itself. GPT-Live-1.1 does the talking and listening, and your backend model does the work. If you want a voice that sounds natural while a stronger model handles the hard parts, that separation is a feature. If you want one model to do everything, it is a limitation worth knowing before you build.
Why Standby Mode Is the Real News
Voice billing has always had an awkward problem. A phone agent or in-app assistant stays connected while the caller reads a document, thinks, or waits on hold, and every one of those silent seconds used to cost the same as talking. Standby fixes that. WebRTC sessions can pause inference during inactivity and resume the moment someone speaks, with idle timeouts configurable from 5 to 60 seconds. The session keeps its connection and ID while paused, so nothing has to reconnect.
The math is simple. An active minute costs $0.05, or $3.00 an hour. A standby minute costs $0.02, or $1.20 an hour. For a support line where callers spend half the call reading an account number off a card, that is real money across thousands of calls a day. Standby applies to both GPT-Live-1 and GPT-Live-1.1, so existing users get the saving without switching models.
There is a catch worth stating plainly. Backend model and tool usage are billed separately, so $0.05 a minute is the floor, not the total. A session that calls an expensive reasoning model on every turn will cost far more than the voice line itself. Anyone budgeting a voice product should price the whole stack, not the headline rate.
Who Should Switch, and Who Should Wait
If you already run GPT-Live-1 in production, the switch is low risk. Same endpoint, same pricing, better pronunciation and accent consistency, and the standby saving applies either way. If you run GPT-Realtime, the choice is architectural: GPT-Live-1.1 is the better voice frontend, but it asks you to bring your own backend brain, which means more moving parts.
The timing is also worth noting. ChatGPT users lose the older Advanced and Standard voice modes on December 11, as OpenAI consolidates everything onto GPT-Live. The company is clearly betting that always-on voice is the next default interface, and standby pricing is the clearest sign of that bet. A voice agent that costs money while nobody is talking cannot stay open all day. One that costs $1.20 an hour while waiting can.
GPT-Live-1.1 is an incremental release with one genuinely useful idea inside it. The voice got better, which is nice. The economics of leaving it running got better, which is the part that changes what people build. Try it on the session your users actually abandon halfway through, the one where they go quiet for thirty seconds to find a document. That is where the saving lives, and where the better accents will either earn their keep or not.


