- Get link
- X
- Other Apps
Featured Post
By Mesum Mukhtar
LearnDailyAI
-
- Get link
- X
- Other Apps
On 19 November 2026, OpenAI put its newest voice model into the hands of developers. GPT Live 1.1 is now available in the OpenAI API, according to the OpenAI API changelog, and it is the same model family that has powered voice in ChatGPT Plus and Pro since 15 October 2026. The headline change is simple: the model listens while it speaks, so a conversation with it feels far closer to talking with a real person.
There is also a price cut hiding in the launch. A new standby mode drops the cost of an idle voice session from 5 cents a minute to 2 cents a minute, which is 60 percent cheaper. Ten idle minutes now cost 20 cents instead of 50 cents, and the WebRTC call stays open the whole time. For anyone building a voice agent that waits on hold, answers phones or sits quietly between questions, that change alone is worth reading about.
Watch the official video
This video is OpenAI's own upload from the official OpenAI YouTube channel, press play and watch without leaving this page. It answers questions about GPT Live, including how it differs from earlier voice models and what developers use it for.
Answering your questions about GPT Live. OpenAI explains the model in its own words.
Watch Answering your questions about GPT Live on YouTube
What GPT Live 1.1 actually does
The core idea is full duplex conversation. Most voice assistants work in turns: you speak, then you wait, then the assistant answers. GPT Live 1.1 listens and speaks at the same time, so you can interrupt it, talk over it or change your mind halfway through a sentence, and it keeps up. That is how human conversation works, and it is the single biggest reason this model feels different from a normal voice bot.
OpenAI lists five clear improvements over GPT Live 1 in the changelog. Here they are, in plain English:
- Better pronunciation: words come out cleaner and clearer, so the voice sounds more natural.
- More consistent accents: the accent stays steady across languages instead of drifting.
- Better transcription: the model hears multilingual voice conversations more accurately.
- Stronger instruction following: it sticks to your instructions more reliably.
- Better handling of precise details: letters, numbers and exact details come through more reliably.
Taken together, these are small upgrades with a large effect. A voice agent that mishears a single letter in an email address or a single digit in a phone number can cause a real problem, so accuracy on small details matters more in voice than it does in text. GPT Live 1.1 is built to close exactly that gap.
Under the hood, streaming and function calling are supported, so the model can call your tools and speak the result without a long pause. Text and audio are supported as inputs and outputs, while images and video are not, so think of it as a voice model with a text side, not a camera.
Hand the hard work to your backend
A voice model is brilliant at talking, but talking is not the whole job. When a caller asks for an order status, a booking change or a long calculation, the voice model should hand that work to your backend and keep the conversation flowing while a backend model of your choice does the deeper work. That is the delegation pattern OpenAI describes in the OpenAI API changelog and the GPT Live 1.1 system card, and it is the right way to think about building with this model.
In practice, the split is simple. Let GPT Live 1.1 own the conversation: greet the caller, ask the next question, confirm what it heard and explain the answer in a friendly voice. Let your backend own the facts: look up the order, check the calendar, run the numbers and return a short answer the voice model can speak. The caller never notices the handover, and your app stays fast because the heavy work happens where it belongs.
Twelve languages and nine new voices
GPT Live 1.1 speaks widely. The launch summary from AlphaSignal lists 12 languages:
- English
- Spanish
- French
- German
- Italian
- Portuguese
- Dutch
- Polish
- Russian
- Japanese
- Korean
- Mandarin Chinese
If you build for a global audience, that list matters. A voice agent that speaks your caller's language, with an accent that stays steady, feels local instead of foreign.
OpenAI also added nine natural voices on 16 October, and each one is tied to a language:
- German: Nuri, Juni and Flitz
- French: Brise and Sillage
- Hindi: Nira
- Japanese: Harema and Shida
- Korean: Noeul
Then there is Prompt to voice, released on the same day, 16 October. Describe the voice you want in plain words, through the API or the voice Playground, and the system builds a custom voice from your description. You no longer need a recording studio or a voice actor to give your app a voice of its own.
GPT Live 1 vs GPT Live 1.1 vs ChatGPT voice
Three names, one family, three different jobs. Here is how they line up:
The history is short but fast. GPT Live 1 launched on 10 September 2026 with the live sessions endpoint /v1/live/sessions, and outbound SIP calling became available to all organisations on 2 November 2026. Seventeen days later, GPT Live 1.1 arrived, which tells you how quickly OpenAI is moving on voice.
In the ChatGPT app, GPT Live 1.1 is the default voice model for Plus and higher tier users, while GPT Live 1.1 Quip is the default for everyone else. Quip is the successor to GPT Live 1 mini, and GPT Live 1.1 is the successor to GPT Live 1, according to the GPT Live 1.1 system card.
| Feature | GPT Live 1 | GPT Live 1.1 | ChatGPT voice |
|---|---|---|---|
| Launch | 10 September 2026 | 19 November 2026 | 15 October 2026 |
| Where it runs | In the OpenAI API | In the OpenAI API | In the ChatGPT app |
| Best for | Developers building voice apps | Developers building voice apps | ChatGPT Plus and higher tier users |
| Accents and pronunciation | The original release | Improved | Same as GPT Live 1.1 |
| Multilingual transcription | The original release | Improved | Same as GPT Live 1.1 |
| Instruction following | The original release | Stronger | Same as GPT Live 1.1 |
| Precise details like letters and numbers | The original release | Better handling | Same as GPT Live 1.1 |
| Standby mode | Yes, for WebRTC sessions | Yes, for WebRTC sessions | Not stated for the app |
The short version is this: if you are building a product, GPT Live 1.1 is the model to use, because it is the newest and the cheapest to keep idle. If you are simply chatting, the voice in ChatGPT already runs on GPT Live 1.1.
Standby mode: the money saver
Standby mode is the most practical part of this release, because it attacks the quiet cost that nobody notices until the bill arrives. A voice agent spends much of its life waiting: waiting for a caller to find a document, waiting on hold with another service, waiting for a human to pick up. Until now, every waiting minute cost the full active rate. Standby changes that.
Here is how it works. Set an idle timeout between 5 and 60 seconds. If nobody speaks for that long, the model pauses inference while the WebRTC call stays open, and the rate drops at once. The moment someone speaks again, the model resumes, and the conversation carries on as if nothing happened. Standby works for both GPT Live 1 and GPT Live 1.1 WebRTC sessions, so even apps on the older model get the saving.
The maths is simple, and it is worth seeing in a table:
| Session state | Per minute | Per hour | 10 idle minutes |
|---|---|---|---|
| Active | $0.05 | $3.00 | $0.50 |
| Standby | $0.02 | $1.20 | $0.20 |
| You save | $0.03 | $1.80 | $0.30 |
The headline price has not changed: voice sessions cost $0.05 a minute while active, which is $3 an hour. Backend model and tool use are billed separately, so budget for those on top of the voice minutes. Standby does not make the active minutes cheaper, it makes the quiet minutes cheaper.
Picture a support line that takes 100 calls a day, with each call idle for 10 minutes while the agent checks something. At the active rate, that idle time costs $50 a day. On standby, it costs $20 a day. That is $30 saved every day, $900 in a 30 day month, from a single setting.
What you can build with it
A voice model that listens while it speaks opens a door that text chat never could: the phone call. Here are four real uses:
- Phone agents: an agent that answers calls, books appointments and takes messages around the clock.
- Customer support: a support voice that listens to the whole problem, asks one clear question at a time and hands the hard cases to a human with a full summary.
- Voice assistants: a hands free assistant for driving, cooking or working, where typing is not an option and the user talks naturally.
- Language practice: a patient tutor that chats in Spanish, French, Japanese or Korean, corrects pronunciation gently and never gets tired of the same question.
To compare voice models with the text models you already know, read our Gemini review and our Claude review.
Limits to know
GPT Live 1.1 is strong, but it has clear limits, and knowing them now saves you a redesign later. The system card is honest about what the model can and cannot do, so here is the list in plain English.
- No structured outputs: the model does not return neat labelled data, so your backend must parse the spoken answer.
- No fine tuning: you cannot train the voice model on your own data, so steer it with prompts and instructions instead.
- Text and audio only: images and video are not supported as inputs or outputs, so this is a voice model, not a vision model.
- Backend costs are separate: the voice minutes are only part of the bill, because your backend model and tools are billed on top.
Voice chats raise privacy questions too, and they deserve a serious answer before you ship. Our post on how Muse showed people the wrong private chats is a reminder that a voice agent can repeat the wrong thing to the wrong person, so keep private data out of the voice layer.
None of these limits is a reason to skip the model. They are reasons to design around it.
Also in the same changelog
One short note for readers who follow model releases: GPT 6.2 Astra and Sol arrived on 17 November for reasoning and coding tasks. That is a separate story, and we will cover it on its own when the details are confirmed.
Verdict: GPT Live 1.1 is a careful upgrade, not a revolution: better accents, better hearing and better manners in conversation, plus a standby mode that cuts the cost of waiting by 60 percent. If you build voice products, move to GPT Live 1.1 and turn on standby. If you only chat, you already have it in ChatGPT, so just talk to it.
Try it yourself
- Listen to the voices: open the OpenAI voice Playground and try Nuri, Brise and Nira, then ask which one sounds most natural to you.
- Do the standby maths: take a voice app you use, guess how many minutes it sits idle in a day, and work out the saving at 3 cents a minute.
- Build a tiny phone agent: sketch the flow for a booking agent on paper first, then mark the moments where the voice model should hand work to the backend.
Quick quiz
1. On what date did GPT Live 1.1 arrive in the OpenAI API?
Answer: 19 November 2026.
2. How much does standby mode cost per minute, and how much does it save against the active rate?
Answer: 2 cents a minute, which saves 3 cents a minute against the 5 cent active rate, or 60 percent.
3. Name one thing GPT Live 1.1 does not support.
Answer: Structured outputs and fine tuning are not supported, and images and video are not supported as inputs or outputs.
Your turn
Would you trust a voice agent to answer your phone, or would you still want a human on the line? Tell us in the comments, share this post with a friend who builds apps, and follow @learningdailyai on Instagram for the next launch, explained simply, every day.
All logos and videos in this post are official OpenAI media, © OpenAI, used here for editorial coverage.



Comments
Post a Comment