Skip to content

AI Tools

What Is GPT-Live-1.1?

Quick Answer

GPT-Live-1.1 is OpenAI’s full-duplex voice model for multilingual conversations, model id gpt-live-1.1. It listens while it speaks, handles the conversation itself, and hands reasoning and tool work to a backend model or agent through delegation. It reached the API on November 19, 2026 through the Live API (v1/live/sessions), after powering ChatGPT voice since October 15, 2026. Compared with GPT-Live-1, OpenAI says it transcribes speech more accurately and keeps a more consistent accent and voice across languages. Voice sessions cost $0.05 per minute while active and $0.02 per minute on standby (WebRTC only), billed per second, with backend model and tool usage billed separately. There are no published benchmarks — OpenAI’s model page shows no performance score.

What Is GPT-Live-1.1?

A normal voice assistant works in turns: you speak, it thinks, it speaks, you wait. GPT-Live-1.1 is full duplex, which means it keeps listening while it is talking — a caller can interrupt, correct themselves or add a detail mid-sentence, and the model adjusts instead of talking over them. That is the whole point of the GPT-Live line: a phone call, not a walkie-talkie.

The design splits the job in two. GPT-Live handles the conversation — listening, speaking, and deciding when to ask the backend for help — while the backend handles delegated tasks: looking up an order, checking a calendar, running tools, doing the reasoning. You describe in plain language when the model should hand a task over, and you choose the backend independently of the voice model. With Responses delegation, OpenAI runs the backend model for you; with client delegation, you connect your own model, agent harness or service. Backend work can keep running even when the caller interrupts, and your application decides whether to finish or cancel it.

GPT-Live-1.1 works with text and audio in and out. Image and video input are not supported, and neither are structured outputs, fine-tuning or predicted outputs — streaming and function calling are supported. It connects over WebRTC (recommended for browser, mobile and server apps), WebSocket (wss://api.openai.com/v1/live/sessions), or SIP for phone calls, with a sideband WebSocket available to monitor or control a session from your backend.

Release Date and Availability

  • October 15, 2026 — OpenAI rolls out GPT-Live-1.1 in ChatGPT, Codex and dots, with 13 new voices and considerably better performance in French, German, Hindi, Mexican Spanish, Korean, Japanese and many other non-English languages. The announcement says it is “coming soon to the API”. The smaller GPT-Live-1.1 Quip rolls out alongside it for Free and Go users.
  • November 19, 2026 — GPT-Live-1.1 is added to the OpenAI API changelog: use gpt-live-1.1 in an existing Live session configuration, on the v1/live/sessions endpoint. Standby configuration arrives the same day for GPT-Live-1 and GPT-Live-1.1 WebRTC sessions.
  • ChatGPT — the full model runs on Plus (3 hours per rolling 24 hours), Pro at $100 a month (15 hours), Pro at $200 a month (unlimited, subject to abuse guardrails), and Business Standard (3 hours) and Business Premium (15 hours), with extra usage at 1.25 credits per minute on Business plans. Free users get limited access to the smaller GPT-Live-1.1 Quip, and Go users get 3 hours of Quip.
  • API — available to API customers through the Live API. GPT-Live-1.1 and GPT-Live-1 have separate concurrent-session rate limits, and each model’s limit applies across inference regions.
  • Predecessor — GPT-Live-1 became generally available in the API on September 10, 2026, at the same $0.05 per minute, billed per second.

The timeline matters: November 19 is the date developers got the model, not the date the model was born. If you have used ChatGPT voice since mid-October, you have already been talking to GPT-Live-1.1.

Key GPT-Live-1.1 Features

1. Full-duplex conversation — it listens while it speaks

The caller and the assistant can speak at the same time. Transcript events arrive separately for input and output speech, so an app can caption both sides independently while they overlap. Interruptions work the way they do on a real call.

2. Better accents, pronunciation and transcription

OpenAI’s stated improvements over GPT-Live-1: it transcribes speech more accurately and maintains a more consistent accent and voice across languages, with better pronunciation. The model page leads with exactly this comparison, and the October announcement names French, German, Hindi, Mexican Spanish, Korean and Japanese as the languages with considerably better performance.

3. Standby mode — idle time at $0.02 a minute

For WebRTC sessions, set session.wake_policy.type to speech and session.idle_timeout_s to a whole number from 5 to 60. After that many seconds without an input or output transcript, inference pauses and the session bills at the $0.02 per minute standby rate instead of $0.05. The microphone track stays active, the connection and session ID stay the same, and the session resumes — with its conversation history restored — the moment someone speaks. A voice activity detector threshold (vad_threshold, 0 to 1, default 0.5) controls how strong the evidence of speech must be to wake it. Two cautions from OpenAI’s own guide: after standby, the first response to a short utterance may take longer, and standby is WebRTC only — it does not apply to WebSocket sessions. Standby arrived on November 19 for GPT-Live-1 as well, so it is a new billing mode for the whole GPT-Live line, not a 1.1 exclusive.

4. Delegation to a backend, in plain language

You tell GPT-Live in plain language when to hand a task to your backend, and the backend does the reasoning and tool use. Responses delegation lets OpenAI run the backend model and pass context and results back and forth; client delegation connects any model, agent harness or service you already run. Delegation is now more natural than in GPT-Live-1, according to the October announcement, and the model can call OpenAI’s Decisions API directly while the user is still speaking. Backend model and tool usage is billed separately from the voice session.

5. Thirteen new voices, plus custom voices

The October rollout added 13 voices: Bubbie in English, Aube and Rivage in French, Sonna and Miko in German, Leher and Neer in Hindi, Himari and Rindō in Japanese, Haetsal and Hanul in Korean, and Tinta and Arrullo in Mexican Spanish. The API’s own voice table adds regional options such as Quartz and Ripple (Australian English), Vesper (British), Willow and Stone (Irish), Bossa and Tempo (Brazilian Portuguese) and Beacon, Delta and Cinder (American English), alongside Nuri, Juni and Flitz (German), Brise and Sillage (French), Nira (Hindi), Harema and Shida (Japanese) and Noeul (Korean). You can also create a custom voice from a text description or an approved recording — creating one is free, and the speech it generates uses regular GPT-Live session billing. The voices you see may vary by region and account.

6. WebRTC, WebSocket, SIP and partner integrations

WebRTC is recommended for audio connections in browser, mobile and server apps — audio travels on media tracks and JSON events on a data channel. WebSockets are a supported alternative, with one connection carrying audio and control events. Outbound SIP calling connects GPT-Live to phone numbers through your SIP provider. If your app already uses LiveKit, Twilio, Telnyx, Bandwidth or Daily/Pipecat, OpenAI documents partner integrations that connect their existing calls and audio streams to GPT-Live.

7. Recordings, forking and usage events

Set store to true when you create a session to make it forkable later, and download the finished recording as a stereo WAV file (input audio in the left channel, output audio in the right). session.usage.updated events report cumulative voice usage in seconds — split into active and paused seconds — plus the latest context-window usage as a ratio. Closing is graceful: send session.close, wait for session.closed, and read the final usage and the reason the session ended.

GPT-Live-1.1 Pricing Explained

Voice sessions are billed by duration, per second. Backend model and tool usage is billed separately, at that model’s own rates.

Model Active rate Standby rate
gpt-live-1.1 $0.05 / minute $0.02 / minute
gpt-live-1 $0.05 / minute $0.02 / minute

A worked example: a 10-minute active call costs $0.50. A 60-minute support session with 20 minutes of talking and 40 minutes on hold costs 20 × $0.05 + 40 × $0.02 = $1.80, against $3.00 if the whole hour were billed active — a saving of $1.20 (40%). Standby only helps when the pauses are real: OpenAI’s guide suggests comparing the avoided voice time against session-creation costs and the slower first reply after standby.

Backend costs are separate and can be larger than the voice bill. For example, a delegated task that uses 5,000 input tokens and 1,000 output tokens of gpt-6.2-sol — $2.00 input and $10.00 output per 1M tokens on OpenAI’s pricing page — adds $0.01 + $0.01 = $0.02 to that session. A heavy reasoning backend on a long call can cost far more than the voice itself, so budget the two lines separately.

In ChatGPT, voice is included in the plan with rolling 24-hour limits:

ChatGPT plan Live model Voice in Chat allowance
Free GPT-Live-1.1 Quip Limited access
Go GPT-Live-1.1 Quip 3 hours
Plus GPT-Live-1.1 3 hours
Pro ($100/month) GPT-Live-1.1 15 hours
Pro ($200/month) GPT-Live-1.1 Unlimited, subject to abuse guardrails
Business Standard GPT-Live-1.1 3 hours, then 1.25 credits per minute
Business Premium GPT-Live-1.1 15 hours, then 1.25 credits per minute

There is no free tier listed for GPT-Live in the API — voice sessions are billed from the first second.

GPT-Live-1.1 vs GPT-Live-1

Spec GPT-Live-1 GPT-Live-1.1
API availability Generally available 10 Sep 2026 Added to the API 19 Nov 2026
Active rate $0.05 / minute $0.05 / minute
Standby rate $0.02 / minute $0.02 / minute
Transcription accuracy Standard More accurate, per OpenAI
Accent and voice consistency Can drift across languages More consistent across languages
Pronunciation Standard Improved
Standby Available (added 19 Nov 2026) Available (added 19 Nov 2026)
Voices API voice set Same API set, plus 13 new ChatGPT voices and prompt-to-voice custom voices
Delegation Responses or client Same, with more natural delegation
Endpoint v1/live/sessions v1/live/sessions
Modalities Text and audio Text and audio
Best for Existing Live apps New builds, multilingual voice, long idle sessions

The upgrade is a model-name swap in your session configuration, not a migration: same endpoint, same price, same delegation modes. Pick 1.1 for new builds and for any app where accents or transcription quality matter; there is no pricing reason to stay on GPT-Live-1.

Performance and Benchmarks

All performance figures here are OpenAI’s own claims from its model page, changelog and guides, unless marked otherwise, and have not been independently verified by faqsnow.

There are no published benchmarks for GPT-Live-1.1 — no accuracy scores, no latency figures, no leaderboard entry. The model page shows no performance score at all, and OpenAI describes the improvements in words (“transcribes speech more accurately”, “more consistent accent and voice”) rather than numbers. That is the honest state of the evidence: the quality claims are the vendor’s, and the only hard numbers are engineering limits, not quality scores. The figures OpenAI does publish are: $0.05 per active minute and $0.02 per standby minute, billed per second; instructions up to 16,384 tokens; startup history up to 128 messages and 8,192 combined tokens; context appends up to 500 tokens each; and context usage reported as a ratio against the model’s limit. OpenAI does not publish the context limit itself as a token count.

How to Use GPT-Live-1.1

1. If you use ChatGPT, just open Voice

In ChatGPT, open Settings, then Voice, and pick a voice — Live is the default voice experience. Plus and Pro users have been talking to GPT-Live-1.1 since October 15; Free and Go users get the smaller Quip version. Check the rolling 24-hour allowance for your plan before a long session.

2. If you build apps, start with the WebRTC quickstart

You need a microphone, a browser page served over HTTPS or localhost, and a trusted server holding an OpenAI project API key — keep the key on the server, never in the browser. Set the session model to gpt-live-1.1, write a short conversation prompt that says when GPT-Live should ask the backend for help, then connect the browser’s microphone and audio playback through the quickstart’s browser and server example. Wait for session.started, then speak — and ask a question that needs current information to see the backend search work.

3. Add delegation before you add features

Choose Responses delegation to let OpenAI run the backend model, or client delegation to connect your own agent or service. Write the delegation rules in plain language — when Live should hand off, what it may decide alone, and what needs the user’s confirmation. Your application still runs its own function tools and enforces permissions and confirmations.

4. Configure standby for always-on sessions

For WebRTC sessions that wait between calls, set session.wake_policy.type to speech and session.idle_timeout_s to a value from 5 to 60. Keep the microphone track active while the session is on standby, and listen for session.paused and session.resumed on the sideband or data channel. If the gaps are long enough that the connection itself is not needed, close the session instead and start a new one when the user returns — a closed session cannot listen for them.

5. Close gracefully and read the final usage

Install the session.closed listener before sending session.close, then wait for the event and read the final usage.seconds, the close reason and the session snapshot. Sending close cancels queued Responses work, so finish any delegated task your application needs first. Track backend token usage separately from the voice seconds.

6. Test with one real conversation, then check the bill

Run a single realistic call, read session.usage.updated for the voice seconds and the backend’s token usage, and compare the split before you scale. A support line that spends most of its time on hold should show most of its minutes at the $0.02 standby rate.

Best Uses

  • Phone support agents — outbound SIP calling plus delegation to a backend that looks up orders and runs tools while the caller keeps talking
  • Always-on voice assistants — in-car helpers and home assistants that wait for speech, where standby turns idle minutes from $0.05 into $0.02
  • Multilingual voice apps — the accent and transcription improvements are the headline change, and the new voices cover French, German, Hindi, Japanese, Korean and Mexican Spanish
  • Hands-free task completion — the caller talks while the backend books, checks or files, and GPT-Live explains the result when it arrives
  • Prototyping a voice product — the WebRTC quickstart plus a Responses backend gives a working voice agent with web search in one example

Limitations and Things to Know

  • No published benchmarks. The quality improvements are OpenAI’s words, not measured scores. Test transcription and accent handling in your own languages before promising them to users.
  • Standby is WebRTC only. WebSocket sessions cannot use it, and the first reply after standby may take longer than a normal reply.
  • No context-window token count is published. OpenAI reports context usage only as a ratio, and says GPT-Live manages context automatically as the conversation grows. Do not quote a token figure for this model — the 128,000-token number in some coverage is not in OpenAI’s documentation.
  • Backend costs are separate and uncapped by the voice rate. A reasoning-heavy backend can cost more than the voice session itself; budget voice seconds and backend tokens as two lines.
  • No image or video input. Text and audio only, and structured outputs, fine-tuning and predicted outputs are not supported.
  • Sessions have a duration limit. A session can end with the reason “expired”, but OpenAI does not publish the limit in these docs — plan for long sessions to be interrupted.
  • Voices vary by region and account. The 13 new voices and the API voice table are what OpenAI lists; your account may show a different set.
  • ChatGPT voice is capped per rolling 24 hours. Plus gets 3 hours and Pro at $100 gets 15 hours; ChatGPT notifies you when you reach the limit.
  • The price is the same as GPT-Live-1. If your app already runs on GPT-Live-1 and does not need the accent or transcription improvements, the upgrade saves nothing — its value is quality and standby, not cost.

The honest read: GPT-Live-1.1 is a quality update with one genuinely useful billing idea. The standby rate matters for always-on voice apps; for everyone else, the change is better accents in more languages at exactly the same price.

Frequently Asked Questions

What is GPT-Live-1.1?

OpenAI’s full-duplex voice model for multilingual conversations. It listens while it speaks, handles the conversation itself, and delegates reasoning and tool work to a backend model or agent. The model id is gpt-live-1.1.

When was GPT-Live-1.1 released?

It rolled out in ChatGPT, Codex and dots on October 15, 2026, and was added to the developer API on November 19, 2026. The November date is the API launch, not the model’s first appearance.

How much does GPT-Live-1.1 cost?

In the API, $0.05 per minute while active and $0.02 per minute on standby (WebRTC only), billed per second. Backend model and tool usage is billed separately. In ChatGPT it is included in the plan, with rolling 24-hour limits — 3 hours on Plus, 15 hours on Pro at $100 a month.

What is standby mode?

A WebRTC billing state that pauses inference after 5 to 60 seconds of silence (you choose the timeout) and bills the idle time at $0.02 a minute instead of $0.05. The microphone stays live and the session resumes with its history when someone speaks. It arrived on November 19, 2026 for GPT-Live-1 as well as GPT-Live-1.1.

What is the difference between GPT-Live-1.1 and GPT-Live-1?

Better transcription accuracy, more consistent accent and voice across languages, and better pronunciation, according to OpenAI. The price, endpoint (v1/live/sessions) and delegation modes are the same, and upgrading is a model-name swap.

Can I use GPT-Live-1.1 for free?

In ChatGPT, Free users get limited access to the smaller GPT-Live-1.1 Quip. In the API there is no free tier listed — voice sessions are billed from the first second.

What is GPT-Live-1.1 Quip?

The smaller version of GPT-Live-1.1, rolled out alongside it on October 15, 2026. It powers Voice on ChatGPT Free and Go plans and replaces the earlier GPT-Live-1 mini. The API pricing page lists only GPT-Live-1.1 and GPT-Live-1.

How do I use GPT-Live-1.1 in my app?

Through the Live API at v1/live/sessions — WebRTC is recommended, WebSocket is a supported alternative. Set the session model to gpt-live-1.1, add a short conversation prompt, and choose Responses or client delegation for the backend.

Does GPT-Live-1.1 support images or video?

No. It works with text and audio only. Image and video input are not supported, and neither are structured outputs, fine-tuning or predicted outputs.

What is the context window of GPT-Live-1.1?

OpenAI does not publish a token count for it. Sessions report context usage as a ratio against the model’s limit, and GPT-Live manages context automatically as the conversation grows. The 128,000-token figure in some coverage is not in OpenAI’s documentation.

Final Verdict

GPT-Live-1.1 is the rare model update where the price list tells the story: nothing changed on the active rate, and the new thing is what happens when nobody is talking. Standby at $0.02 a minute makes always-on voice affordable for the first time on OpenAI’s platform, and the accent and transcription improvements are real for anyone building in French, German, Hindi, Japanese, Korean or Mexican Spanish. The honest limits are the missing numbers — no benchmarks, no published context size, no session-length figure — so treat the quality claims as the vendor’s until you test them in your own language. If you already talk to ChatGPT by voice, you have been using this model since October 15. If you build voice apps, swap the model name, wire up standby, and watch the idle minutes — that is where this release pays.

Editorial update: November 21, 2026. This article is based on OpenAI’s API changelog (19 Nov 2026), the GPT-Live-1.1 model page, OpenAI’s API pricing page, the “Getting started with GPT-Live” and “Managing GPT-Live sessions” guides, OpenAI’s developer community announcement (15 Oct 2026) and the ChatGPT Voice help page. Performance claims are OpenAI’s own unless marked otherwise, and have not been independently verified by faqsnow.