Skip to content
, ,

GPT-Live-1.1 API Guide: Standby Mode at $0.02/Min, Pricing and Setup

9–14 minutes
Woman in a podcast studio holding a phone with the OpenAI logo, hook $0.02 idle GPT-Live-1.1 API

OpenAI has put GPT-Live-1.1 into its API. The model ID gpt-live-1.1 appeared in the OpenAI developer changelog on 19 November 2026, alongside a new Standby setting that cuts the price of idle voice sessions from $0.05 to $0.02 per minute.

GPT-Live-1.1 is the same voice model family that has powered ChatGPT voice since 15 October. Developers now get its better accents, cleaner transcription and more reliable handling of spelled letters and numbers in their own apps.

This guide covers what changed, the full price list with worked cost examples, how to switch from GPT-Live-1, how Standby works, a Python starter, the voice list, limits, errors and how it compares with gpt-realtime-2.1.

TL;DR

  • What: gpt-live-1.1, a full-duplex voice model (it listens while it speaks) on the Live API endpoint v1/live/sessions.
  • When: OpenAI changelog entry dated 19 November 2026.
  • Better than GPT-Live-1 at: pronunciation, accent consistency across languages, transcription, letters, phone numbers and order codes (OpenAI’s claims; no benchmark numbers published).
  • Price: $0.05 per minute active, $0.02 per minute in Standby, billed per second. Backend model and tool use are billed separately.
  • Standby: WebRTC only. Set a wake policy of speech and an idle timeout of 5-60 seconds; the call stays open at the lower rate until someone speaks.
  • Switching: change the model string from gpt-live-1 to gpt-live-1.1 in your Live session config. It does not work on /v1/realtime.

What Is GPT-Live-1.1 and What Changed?

GPT-Live is OpenAI’s newer voice line, separate from the older Realtime models. It is full-duplex: it can hear you while it is talking, so interruptions feel natural. It can also hand harder work to a backend model or to your own code, which OpenAI calls delegation.

GPT-Live-1 arrived in the API on 10 September 2026. On 15 October, OpenAI moved ChatGPT voice onto the GPT-Live family, with GPT-Live-1.1 as the default for Plus and higher plans. We explain the ChatGPT side in our ChatGPT Voice Mode guide.

The 19 November changelog brings that 1.1 model to developers. OpenAI’s own wording is that GPT-Live-1.1 “improves pronunciation, accent consistency, and transcription in multilingual voice conversations.”

The three improvements, in plain words

  • Accents stay put. Coverage of the launch notes the model is less likely to slide into an English accent while speaking German, Japanese or other languages.
  • Better hearing. The model page says it transcribes speech more accurately than GPT-Live-1.
  • Letters and numbers. OpenAI says spelled names, phone numbers, order IDs and confirmation codes are handled more reliably. That is the classic failure point for support bots.

Note: OpenAI has not published benchmark scores, word error rates or latency figures for 1.1. All quality claims in this article are vendor-reported.

Release timeline

Date (2026) What happened
10 Sep GPT-Live-1 launches in the API (Live endpoint)
15 Oct ChatGPT voice moves to GPT-Live; 1.1 is the default for Plus and above, a lighter “Quip” version for other users
2 Nov Outbound SIP calling opens to all organisations on the Live API
19 Nov gpt-live-1.1 added to the API, plus Standby for GPT-Live-1 and 1.1 WebRTC sessions

Specs at a Glance

Item GPT-Live-1.1
Model ID / snapshot gpt-live-1.1 (only snapshot)
Endpoint v1/live/sessions only (not Realtime, Chat Completions or Responses)
Input Audio, text
Output Audio, text
Image / video input Not supported
Function calling Yes
Streaming Yes
Structured outputs, fine-tuning, predicted outputs Not supported
Transports WebRTC (recommended), WebSocket, SIP
Context window 128,000 tokens by default (Live guide), shared by instructions, text and audio
Instructions limit Up to 16,384 tokens
Default voice marin
Rate limits Concurrent sessions, separate from GPT-Live-1; see your org limits page

OpenAI does not list a knowledge cutoff for 1.1. GPT-Live-1’s page gives 31 July 2025.

Pricing and Real Cost Examples

GPT-Live is billed by session time, not by tokens. That makes it much easier to predict than the token-priced Realtime models.

Item Price Notes
GPT-Live-1.1 active $0.05 / min Billed per second, not rounded up
GPT-Live-1.1 Standby $0.02 / min WebRTC only, while inference is paused
GPT-Live-1 (active / Standby) $0.05 / $0.02 per min Same as 1.1
WebRTC session start 15 seconds billed Credited against duration once the session runs
Backend model + tools Separate Normal Responses API token prices
gpt-live-transcribe (for comparison) $0.017 / min Speech-to-text only
gpt-realtime-translate (for comparison) $0.034 / min Live translation

The Standby cut is 60% per idle minute: ($0.05 minus $0.02) divided by $0.05. Your total saving depends on how much of the call is silence.

Worked examples (voice time only, excluding backend tokens)

Scenario Without Standby With Standby Saving
6-minute support call, talking the whole time $0.30 $0.30 0%
30-min tutoring session, 12 min of talk $1.50 about $0.96 about 36%
One idle hour on an always-on assistant $3.00 $1.20 60%
8-hour kiosk, 45 min of talk $24.00 about $11 about 54%

The tutoring row is 12 × $0.05 + 18 × $0.02. Real numbers will be a little higher, because each pause only starts after your idle timeout (5-60 seconds) runs out.

Tip: Muting the microphone does not stop billing. Only Standby (or closing the session) drops the rate.

How Standby Mode Works

Before Standby, an open GPT-Live call cost $0.05 a minute even in total silence. Many developers said that killed always-on assistants, kiosks and “leave it running” tutors. Standby fixes that without hanging up the call.

  1. You create a WebRTC session with a wake policy of type speech.
  2. The model wakes when it hears speech. You get a session.resumed event.
  3. Silence starts the clock. If there is no input or output transcript for your idle timeout (5-60 seconds), the session pauses and sends session.paused.
  4. The call stays open. Same connection, same session ID, charged at $0.02 a minute.
  5. Someone speaks and inference resumes at the active rate.

The settings

Field (as written in OpenAI’s guide) Value What it does
session.wake_policy.type speech Turns Standby on (set when creating the WebRTC session)
session.idle_timeout_s Whole number 5-60 Seconds of silence before pausing. If left out, the session stays active after the first wake
session.wake_policy.vad_threshold 0-1, default 0.5 How loud or clear speech must be to wake the model
Watch out: keep the mic track live during Standby, or the model cannot hear the wake word. OpenAI also warns the first reply after waking may take a little longer.

Standby does not apply to WebSocket or SIP sessions in the 19 November release. It also does not help during long backend jobs, because the session is not silent while the model is working.

How to Start a GPT-Live-1.1 Session (Python)

The recommended setup is a small server that holds your API key, plus a browser page that captures the mic over WebRTC. If you are new to the OpenAI SDK, start with our OpenAI API Python tutorial first.

  1. Get an API key from platform.openai.com and set it as OPENAI_API_KEY on your server. Never put it in browser code.
  2. Update the SDK: pip install -U openai. The Live examples call client.live.create(...).
  3. In the browser, ask for mic access from a button click (HTTPS or localhost only), create an RTCPeerConnection, add the mic track, open a data channel named oai-events, then create an SDP offer.
  4. POST the offer to your server. The server sends it to OpenAI and returns the SDP answer.
  5. Apply the answer with setRemoteDescription and wait for session.started on the data channel. Do not send a start event yourself.

A minimal server handler looks like this:

from openai import OpenAI, APIError

client = OpenAI()

def create_live_session(offer_sdp: str):
    session = {
        "model": "gpt-live-1.1",
        "instructions": "You are a friendly support agent. Keep answers short.",
        "audio": {"output": {"voice": "marin"}},
    }
    try:
        result = client.live.create(
            session=session,
            transport={"type": "webrtc", "sdp": offer_sdp},
        )
    except APIError as err:
        return {"error": str(err)}, getattr(err, "status_code", 502) or 502
    # result.transport.sdp is the answer for the browser
    return result.model_dump_json(), 201

A successful call returns HTTP 201 with a session ID (treat it as opaque) and the SDP answer. Add the Standby fields from the table above to the session when you create it.

Note: OpenAI’s guide lists the Standby fields as dotted paths and does not show a full JSON sample. Check the “Configure standby” section of the Live conversations guide for the exact nesting before you ship.

Moving from GPT-Live-1

  • Change "model": "gpt-live-1" to "gpt-live-1.1". Prices are identical.
  • Concurrency limits are counted separately for each model, so check your 1.1 limit in organisation settings before a big rollout.
  • You cannot switch model mid-call. Changing model, instructions, input, audio or store needs a new session.

Ending a call properly

Send session.close on the data channel, wait for session.closed, then close the peer connection and stop the mic tracks. The closed event carries final usage, so skipping it leaves your cost tracking incomplete.

Voices and Languages

Voices are set with audio.output.voice (default marin). The Live guide lists regional voices, and custom voices are possible.

Language / accent Voices
English (Australian) quartz, ripple
English (British / Irish) vesper; willow, stone
English (North American) gleam, meridian
English (Filipino / Southern US) beacon; delta, cinder
Portuguese (Brazil) bossa, tempo
German nuri, juni, flitz
French brise, sillage
Hindi nira
Japanese harema, shida
Korean noeul

For Indian users, the Hindi voice nira plus 1.1’s better accent handling is the most interesting part of this update.

GPT-Live-1.1 vs GPT-Live-1 vs gpt-realtime-2.1

GPT-Live-1.1 GPT-Live-1 gpt-realtime-2.1
Endpoint /v1/live/sessions /v1/live/sessions /v1/realtime
Billing $0.05/min, $0.02 Standby $0.05/min, $0.02 Standby Tokens: audio $32 in / $64 out per 1M
Full-duplex (listens while speaking) Yes Yes Not described this way
Image input No No Yes
Multilingual accents Improved (vendor claim) Baseline n/a
Best for New voice agents, phone bots, tutors Existing apps (switch to 1.1) Apps already on Realtime, vision + voice

If you already run a Realtime agent, our OpenAI Realtime API guide covers that side. Note that older gpt-realtime and gpt-realtime-mini models shut down on 20 January 2027, per OpenAI’s deprecations page.

Limits, Privacy and What Is Still Unknown

  • Cloud only. There is no self-hosted or on-premises version.
  • No free tier. GPT-Live-1 lists 50 concurrent sessions on Build, 300 on Launch and 500 on Grow; 1.1 tier numbers are only shown in your org settings.
  • Recordings. With store on, stereo WAV recordings are kept for 30 days. store defaults to false; Zero Data Retention forces it off and disables forking.
  • Safety cut-offs. Some moderation events end the session (reason content); others just stop the assistant’s audio and send an error.
  • Not published: benchmark scores, latency figures, a 1.1 knowledge cutoff, and whether gateways such as Vercel AI Gateway (listed at $3.00 per session hour) pass through the Standby rate.

Troubleshooting Common Errors

Problem Likely cause Fix
Model not supported error Using gpt-live-1.1 on /v1/realtime Call v1/live/sessions (client.live.create)
immutable_field_update Trying to change delegation type, model or voice mid-call Start a new session with the new settings
No audio heard Browser blocked autoplay Show a play button; start audio from a user click
“Timed out while gathering ICE candidates” Network or firewall blocks WebRTC Test another network; allow UDP; check TURN setup
Bill higher than expected in quiet periods Standby not enabled, or only muting the mic Set wake policy speech plus an idle timeout
Model never wakes from Standby Mic track stopped, or VAD threshold too high Keep the track live; try vad_threshold 0.3-0.5
Session closed with reason expired Hit the session duration limit Reconnect (note the 15-second start charge)
429 or concurrency error Too many live sessions for your tier Check org limits; see our OpenAI API errors guide

My Take

The model upgrade is welcome, but Standby is the real headline. Per-minute billing made GPT-Live simple to price, yet it punished any app where people pause, think or walk away. Cutting idle time to two cents a minute makes always-on assistants, classroom tutors and shop kiosks viable.

The 1.1 accuracy gains matter most for phone and support bots, where one wrong digit in an order number ruins the call. I would still test with your own accents and numbers, because OpenAI has shared no figures.

If you are on GPT-Live-1, switching is a one-line change at the same price. If you are starting fresh on voice, I would pick GPT-Live-1.1 over the token-priced Realtime models unless you need image input.

FAQ

When was GPT-Live-1.1 released in the API?

OpenAI’s developer changelog lists it on 19 November 2026. It had already been the default ChatGPT voice model for Plus and higher plans since 15 October.

How much does GPT-Live-1.1 cost?

$0.05 per minute while active and $0.02 per minute in Standby, billed per second. Any backend model or tool calls are billed separately at normal token rates.

Is GPT-Live-1.1 more expensive than GPT-Live-1?

No. Both models have the same active and Standby prices.

Does Standby work over WebSocket or SIP?

Not in the 19 November release. OpenAI says Standby applies to GPT-Live-1 and 1.1 WebRTC sessions.

Can I use gpt-live-1.1 with the Realtime API?

No. It only works on the Live endpoint, v1/live/sessions. Realtime apps use gpt-realtime-2.1 or 2.1-mini.

Does GPT-Live-1.1 support Hindi?

Yes. The voice list includes a Hindi voice, nira, and 1.1 is tuned for more consistent accents across languages.

Is there a free tier for GPT-Live?

No. GPT-Live-1’s model page marks the free tier as unsupported. You need a paid API organisation.

Does muting the mic save money?

No. A connected session is billed while muted. Use Standby or close the session to stop paying the active rate.

Sources

  • OpenAI developer changelog, 19 Nov 2026 entries for gpt-live-1.1 and Standby (developers.openai.com/api/docs/changelog)
  • OpenAI model pages: gpt-live-1.1 and gpt-live-1 (developers.openai.com/api/docs/models)
  • OpenAI Live guides: “Live conversations” (Configure standby, voices, events) and “Voice over WebRTC”
  • OpenAI API pricing page and deprecations page (developers.openai.com)
  • @OpenAIDevs post on Standby, via Digg AI roundup, 19 Nov 2026
  • AlphaSignal, “OpenAI Ships GPT-Live-1.1 to Developers With 60% Cheaper Idle Sessions”, 19 Nov 2026
  • Data Studios, “OpenAI GPT-Live-1.1 API: Standby Mode at $0.02 per Minute”, 19-20 Nov 2026 (cost scenarios)
  • Vercel AI Gateway model listing: openai/gpt-live-1.1
Free weekly email

Get new AI guides every week

One short weekly email with our newest local AI, ComfyUI and AI tool guides. No spam, no pop-ups, just the good stuff.

Free. Unsubscribe anytime in one click. We only use your email to send this newsletter. See our privacy policy.

Woman in a podcast studio holding a phone with the OpenAI logo, hook $0.02 idle GPT-Live-1.1 API