DEV Community

Cover image for OpenAI Just Put GPT-Live-1.1 in the API, and Idle Voice Sessions Are Now 60% Cheaper
Paul Lovelace
Paul Lovelace

Posted on Originally published at toolsdig.com

OpenAI Just Put GPT-Live-1.1 in the API, and Idle Voice Sessions Are Now 60% Cheaper

Voice agents have had an expensive habit: they charge you full price while nobody is talking. A customer puts the phone down for a minute, the agent sits there listening to silence, and the meter keeps running at the active rate. OpenAI's answer, announced 19 November 2026 and in the API from 20 November, is twofold: a better model, GPT-Live-1.1, and a Standby mode that cuts the idle rate by 60%.

I have not run this hands-on yet — this is a launch-day piece, built from OpenAI's announcement, the model docs and the pricing page — so take the setup details as per the docs, not as my war story. But the pricing change is worth doing the maths on, because it changes which voice products are worth building.

What actually shipped

GPT-Live-1.1 (model id gpt-live-1.1) is now available in the OpenAI API. It is the same model that has powered voice in ChatGPT Plus and Pro since 15 October 2026. GPT-Live-1, the first version, launched on 10 September 2026.

This is a full-duplex voice API: the model listens while it speaks, so interruptions work the way they do in a real conversation. You start sessions through POST /v1/live/sessions, over WebRTC, WebSocket or SIP.

One thing to get straight before you touch any code: this is NOT the Realtime API. GPT-Live lives on /v1/live/sessions, not /v1/realtime, and a Realtime integration is not automatically compatible. If you built on Realtime, this is a migration, not a model swap.

What is better in 1.1

Compared with GPT-Live-1, OpenAI says 1.1 brings:

  • More accurate transcription
  • More natural and consistent accents across languages
  • Better at catching precise details — letters, numbers, the things a voice agent usually gets wrong
  • Better instruction following, and tasks completed more reliably

That third one is the one I would actually test first. Voice agents fail in boring ways: a postcode read back wrong, a digit dropped from an order number. If 1.1 genuinely holds onto precise details better, that is worth more than a nicer accent.

Standby mode is the real story

Here is the change that moves money. Standby mode lets you set an idle timeout of 5 to 60 seconds. After that much silence, the model pauses inference while the WebRTC call stays open. Billing drops from $0.05 to $0.02 per active/idle minute — a 60% cut. The moment someone speaks, the session resumes.

Ten idle minutes now cost $0.20 instead of $0.50. Standby is available for both GPT-Live-1 and GPT-Live-1.1, so you do not need to move to 1.1 to get the cheaper idle rate.

A quick example, because the headline number hides the real saving. Take a voice agent on an eight-hour shift where the caller is actually talking 30% of the time and the line is quiet the other 70%:

  • Old world: 480 minutes at $0.05 = $24.00 a day.
  • New world: 144 active minutes at $0.05 ($7.20) plus 336 standby minutes at $0.02 ($6.72) = $13.92 a day.

That is roughly 40% off the daily bill for a realistic support agent, before you count anything else. The exact saving depends on how much of your session is silence — a receptionist agent that waits between callers saves more than one that talks non-stop.

Two honest footnotes on the pricing. First, billing is per second, never rounded up to a whole minute, so a 40-second call costs 40 seconds. Second, standby applies only to WebRTC sessions while inference is paused — it is not a generic silence discount — and the backend model and tool usage are billed separately. The voice rate is only part of the bill: if your agent delegates reasoning to a backend model, that usage is charged on top.

The session can hand work to a backend model

GPT-Live-1.1 also leans into delegation. In the session config you describe, in plain language, when the model should hand reasoning or tool use to a backend model of your choice — a Responses model, the Decisions API, or your own agent — without breaking the conversation. The voice model stays the voice; the heavy thinking happens elsewhere and the caller never notices the handoff.

That is the right shape for a production voice agent. You do not want the model doing the talking also doing the database lookups and the policy checks. You want it to know when to ask the backend, and to keep the conversation flowing while it does.

The limits, stated plainly

  • Context: 128,000-token window; up to 16,384 tokens of startup instructions; up to 128 prior messages totalling 8,192 tokens. When usage passes 90%, the API automatically spins up a summarised replacement voice engine in the background.
  • Supported: streaming and function calling.
  • Not supported: structured outputs and fine-tuning. That matters. If your agent fills in forms or returns machine-readable data, you cannot force a JSON schema out of the voice model itself — you will need the backend model to do that part, which is exactly what the delegation feature is for.
  • Inputs and outputs: text and audio only. No images, no video.
  • Realtime is a different API. I will say it again because it is the mistake most people will make first: do not point your Realtime code at gpt-live-1.1 and expect it to work.

What I would actually do

If you are already running GPT-Live-1 in production, the move is small: change the model id, set a standby timeout that matches how your callers actually behave (start at 30 seconds and watch the logs), and check the next invoice. The standby saving needs no code beyond the timeout.

If you are on the Realtime API and wondering whether to migrate, do not rush it. GPT-Live-1.1 is the better voice and the cheaper idle session, but the migration is real work, and the lack of structured outputs means your form-filling flows need the delegation pattern. Prototype one flow first — a booking or a support triage — and measure the actual active-to-standby ratio before you commit.

And if you are starting a voice agent from scratch today, start here rather than on Realtime. The model is newer, the idle pricing is built in, and the delegation design is the one you would end up building yourself anyway.

Are you running a voice agent in production — and how much of your session time is actually silence? That ratio decides whether Standby mode is a footnote or the whole story.


Sources: OpenAI announcement via @OpenAIDevs, 19 November 2026; OpenAI API pricing page (GPT-Live: $0.05 active / $0.02 standby per minute, billed per second); OpenAI GPT-Live-1.1 model documentation; independent launch coverage, 20 November 2026. GPT-Live-1 launched 10 September 2026; GPT-Live-1.1 has powered ChatGPT voice since 15 October 2026. Standby applies only to WebRTC sessions while inference is paused. No hands-on testing was performed for this article. Originally published on ToolsDig.

Top comments (2)

Collapse
 
paragonaimag profile image
Paragon •

What stood out for me was the standby mode, not the model upgrade — your $24-a-day example dropping to $13.92 tells the story better than any spec list. Charging full price while nobody's talking always felt like a quiet tax on voice agents, and it's the kind of cost that only shows up once you're running something real.

We're on the other side of voice, running a small podcast that gets a median of 5 downloads per episode, so our scale is tiny — but the silence point lands the same way for us. Most of a conversation is pauses, and any tool that stops charging for the pauses is one built for people who aren't running a call center.

Collapse
 
paullovelace profile image
Paul Lovelace •

Ha — 'quiet tax' is exactly the right phrase for it, that's what it felt like running the numbers. The $24 → $13.92 bit surprised me too, because the model didn't get cheaper, the silence just stopped being billed. And honestly your podcast is the better test case than mine — 5 downloads an episode and the pauses still matter, because most of a real conversation is people thinking, not talking.