Picture a small business owner who has just replaced her front-desk phone with an AI receptionist. The phone rings, the voice agent greets the caller, and then the caller pauses to look up an order number.
Nobody is speaking. The line is open, the agent is listening, and the meter is still running. For the past two months, every one of those silent seconds has cost the same as a second of talking. OpenAI has just changed that maths.
On November 20, OpenAI made GPT-Live-1.1 available in its API, and it arrived with a new standby mode. Idle voice time now drops from $0.05 to $0.02 a minute, a 60% cut.
The model itself is the same voice engine that has powered ChatGPT Plus and Pro since October 15. What is new is that developers can now build with it, and that silence no longer costs full price.
What GPT-Live-1.1 Actually Is
GPT-Live-1.1 is a full-duplex voice model, which means it listens while it speaks. A caller can interrupt it mid-sentence, the way people interrupt each other in a real conversation, instead of waiting for a turn to end.
It takes audio and text in, and it gives audio and text back, streaming the reply as it goes. The model is built for multilingual conversation, and OpenAI describes it as speaking with more natural and consistent accents across languages than its predecessor.
It also transcribes speech more accurately and handles precise details, like letters and numbers, more reliably. That matters when a voice agent is taking down a booking reference or a street address.
There is a second half to the design that is easy to miss. When a request needs real work, GPT-Live-1.1 hands the job to a backend model chosen by the developer. It keeps talking to the caller while that backend runs, then explains the result out loud.
Standby Mode: Paying Less for the Quiet Parts
Standby is the headline feature, and it is aimed at a problem every voice developer knows. Real conversations are full of gaps. People pause to think, to read a document, or to walk to another room.
Until now, an open voice session billed every one of those seconds at the full active rate, whether anyone was talking or not. With standby, a session that goes quiet drops to $0.02 a minute.
The WebRTC connection stays open, the session ID stays the same, and the conversation history comes back when the session resumes. Speech wakes it up again, and the caller never knows the model was resting.
That changes the economics of always-on voice products. An hour of mostly idle connection used to cost $3.00 at the active rate. With standby, the same hour costs $1.20, apart from the few seconds of active time before the idle timeout kicks in.
How Standby Works Under the Hood
Developers switch standby on with two settings when they create a session. The first, a wake policy set to speech, tells the session to wait for voice activity before it starts inference.
The second, an idle timeout of between 5 and 60 seconds, decides how long the model stays active after the last transcript on either side. Once that timeout passes with no input or output transcript, OpenAI pauses inference and the session drops into standby.
The connection itself never closes. When someone speaks again, inference resumes and the conversation picks up where it left off, history intact.
There is one important limit. Standby works over WebRTC only. Sessions running over WebSocket or SIP bill every connected second at the active rate, so developers using those transports get no discount for quiet time.
What GPT-Live-1.1 Improves Over GPT-Live-1
GPT-Live-1 reached general availability on September 10, 2026, at the same $0.05-a-minute active rate. GPT-Live-1.1 keeps that rate, the same endpoint, and the same feature set.
The changes are in voice quality and task handling, not in the price of talking. According to OpenAI Developers, the new model completes tasks more reliably and follows instructions more closely. It delegates complex workflows to the backend model of choice without breaking the flow of conversation.
It also holds a steadier accent across languages, which matters for companies serving customers in more than one country.
OpenAI has not published benchmark numbers for the transcription or reliability gains, so those improvements should be treated as vendor claims until developers measure them in their own products. The standby pricing, by contrast, is a hard number on the pricing page, and it applies to GPT-Live-1 as well as to the new model.
The Maths: What an Hour of Talking Really Costs
Voice sessions are billed by duration, not by tokens, and both rates are charged per second with no rounding up to the next whole minute. Active session time costs $0.05 a minute, or $3.00 an hour. Standby time costs $0.02 a minute, or $1.20 an hour.
The backend that performs delegated work is billed separately, at the normal rates of whatever model or tools the developer configures. That is the part of the bill that does the heavy reasoning, and it is where the real cost of a complex call lives.
Compare that with the token-based Realtime models, which charge $32 per million audio tokens in and $64 per million out. For long, mostly quiet sessions, a per-minute standby rate is far easier to predict.
Who This Matters Most For
The biggest winners are the products that keep a voice session open for a long time. A customer-support agent that waits while a caller checks an account, a receptionist that sits through a hold, or a companion app that listens in the background all spend most of their minutes in silence.
Standby turns those minutes from a cost into a rounding error. It also matters for the small teams building voice products on a budget. A startup running a voice agent for a few hundred users a day can now leave sessions open without watching the meter climb during every pause.
There is a wider context here too. OpenAI is retiring its older Advanced and Standard Voice modes in ChatGPT on December 11, 2026, and moving everyone to GPT-Live. The retirement of the old voice modes means GPT-Live is now the single voice engine across the company, in the app and in the API alike.
The Catches Worth Knowing
Standby is not a free lunch. The WebRTC-only limit means teams on WebSocket or SIP cannot use it, and the slower first reply after waking up could be noticeable in a fast-moving conversation.
Developers will need to tune the idle timeout carefully. Too short, and the session keeps dropping in and out of standby. Too long, and the savings shrink.
There is also a design question that comes with any voice agent. A session that stays open and listening, even at a lower rate, is still a session that is listening.
OpenAI’s model page lists audio and text input and output, streaming and function calling, but no structured output. Developers who need strict formats will have to handle that in the backend.
And while the voice is cheaper, the intelligence behind it is not. A voice agent that delegates every question to a frontier reasoning model will still rack up backend costs. Standby fixes the price of silence, not the price of thinking.
What Happens Next for GPT-Live-1.1
The model is available now through the Live, Realtime, Chat Completions and Responses endpoints, at the model ID gpt-live-1.1, so nothing about how developers connect changes. The announcement, shared by OpenAI Developers, frames it as a straightforward upgrade: same price to talk, cheaper to wait, better at the job in between.
Watch what developers do with the idle time. If standby makes always-on voice cheap enough, the next wave of voice products may be quiet companions that sit in the background of a workday, waiting to be useful.
OpenAI’s own ChatGPT Space, which recently opened to Plus subscribers, shows the company is already thinking in that direction, turning ChatGPT from a chat window into a shared workspace where work lives.
The voice-agent race is no longer just about who sounds most human. It is about who can afford to listen the longest. With GPT-Live-1.1, OpenAI has just made listening a lot cheaper.
The companies building voice products will feel the difference in their very next invoice. As with any powerful AI system, the controls around it matter too, which is why OpenAI built a Control Room kill switch that can halt a rogue model in 35 seconds.



