> ## Content Index
> Fetch the complete content index at: https://signal-engine.ghost.io/llms.txt
> Use this file to discover other available public pages before exploring further.

# OpenAI puts GPT-Live-1.1 in its API and adds a cheaper standby rate for quiet moments on voice calls. What changed, what it costs, and what it means for businesses building voice agents
- URL: https://signal-engine.ghost.io/openai-gpt-live-1-1-api-standby-pricing-voice-agents-what-it-means/
- Published: 2026-11-20T05:10:25.000Z
- Updated: 2026-11-20T05:10:25.000Z
- Description: OpenAI has made its GPT-Live-1.1 voice model available to developers and added standby, which pauses a voice session during silence and bills it at $0.02 a minute instead of $0.05. Here is what changed, what it costs in rands, and what it means for business.
- Author: John Caraher
- Tags: Issue 3, AI, Big Tech, Small Business, #Import 2026-11-20 05:12

Updated 20 Nov 2026, 07:10 SAST

**The short version:** OpenAI has made GPT-Live-1.1, its newest voice model, available to developers through its application programming interface (API), the service businesses use to build OpenAI's models into their own apps[\[2\]](#cite-note-2)[\[3\]](#cite-note-3). OpenAI's developer changelog dates the release 19 November[\[2\]](#cite-note-2). OpenAI says the model transcribes speech more accurately and keeps a more consistent accent and voice across languages than the earlier GPT-Live-1[\[3\]](#cite-note-3). Alongside it comes a new setting called standby. When nobody has spoken for a set time, between 5 and 60 seconds, the model pauses while the call stays connected, and it resumes when someone speaks again[\[1\]](#cite-note-1)[\[4\]](#cite-note-4). Standby time is billed at $0.02 a minute, compared with $0.05 a minute for active time[\[3\]](#cite-note-3)[\[6\]](#cite-note-6) (see [what it costs](#what-it-costs)). Both GPT-Live models now have the same prices[\[6\]](#cite-note-6). Our [analysis](#why-it-matters-for-your-business-analysis) looks at what this means for a South African business thinking about a voice assistant for bookings, orders or customer queries.

Contents

1. [What was announced](#what-was-announced)  
  1. [A new voice model for developers](#new-voice-model)
  2. [Where the announcement appeared](#where-announced)
  3. [How it fits with the rest of GPT-Live](#how-it-fits)
2. [Standby, explained](#standby-explained)  
  1. [What it does](#what-standby-does)
  2. [The catches](#the-catches)
3. [What it costs](#what-it-costs)  
  1. [The rates](#the-rates)
  2. [What counts as active time](#what-counts-as-active)
  3. [What that means in rands](#in-rands)
4. [Other details for developers](#developer-details)  
  1. [Voices and languages](#voices-and-languages)
  2. [Instructions and limits](#instructions-and-limits)
5. [What we could not verify](#what-we-could-not-verify)
6. [Why it matters for your business (analysis)](#why-it-matters-for-your-business-analysis)  
  1. [Voice assistants are getting cheaper to leave running](#cheaper-to-leave-running)
  2. [Work out your own numbers before you build](#work-out-your-numbers)
  3. [Test the accent and the language](#test-accent-and-language)
  4. [Keep a person in the loop](#keep-a-person)
  5. [Mind the waiting time](#mind-the-waiting)
  6. [What to do this week](#what-to-do)
7. [Sources](#sources)

## What was announced

### A new voice model for developers

GPT-Live is OpenAI's family of voice models for live spoken conversations. OpenAI describes GPT-Live-1.1 as a ["full-duplex voice model for multilingual conversations"](https://developers.openai.com/api/docs/models/gpt-live-1.1?ref=signal-engine.ghost.io)[\[3\]](#cite-note-3). Full duplex means it "can listen while it speaks"[\[3\]](#cite-note-3), so a caller can interrupt or add a detail without waiting for the assistant to finish, as in a normal phone conversation.

The model page lists three improvements over GPT-Live-1\. OpenAI says GPT-Live-1.1 "transcribes speech more accurately and maintains a more consistent accent and voice across languages"[\[3\]](#cite-note-3). The [developer changelog](https://developers.openai.com/api/docs/changelog?ref=signal-engine.ghost.io) entry dated 19 November says it "improves pronunciation, accent consistency, and transcription in multilingual voice conversations"[\[2\]](#cite-note-2).

Developers use it by setting the model name `gpt-live-1.1` in a session[\[1\]](#cite-note-1)[\[3\]](#cite-note-3). The changelog says it can be used "in an existing Live session configuration"[\[2\]](#cite-note-2), which suggests developers already using GPT-Live-1 can switch without rebuilding their app. We have not tested this.

### Where the announcement appeared

The news was posted in the Announcements section of the [OpenAI Developer Community forum](https://community.openai.com/t/gpt-live-1-1-has-arrived-in-the-api/1418537?ref=signal-engine.ghost.io) under the username "sps"[\[1\]](#cite-note-1). The page timestamp shows it went up at 04:12 UTC on 20 November, which is 06:12 SAST; we converted the time from the page data. The forum lists "sps" as a community "Leader"; we could not confirm whether the poster works for OpenAI. The main facts in the post (the model name, the standby timeout range and the two prices) also appear on OpenAI's own documentation pages[\[2\]](#cite-note-2)[\[3\]](#cite-note-3)[\[4\]](#cite-note-4)[\[6\]](#cite-note-6), so we rely on those where possible.

The forum post adds some claims that we did not find on the documentation pages we read. It says GPT-Live-1.1 "has powered voice in ChatGPT Plus and Pro since October 15"[\[1\]](#cite-note-1). It also says the model can "understand precise details like letters and numbers, follow instructions, and complete tasks more reliably"[\[1\]](#cite-note-1). These are claims made in the announcement post, and we have not seen independent tests of them.

We did not find an OpenAI blog post, press release or news coverage of GPT-Live-1.1 at the time of writing.

### How it fits with the rest of GPT-Live

OpenAI's [GPT-Live guide](https://developers.openai.com/api/docs/guides/live?ref=signal-engine.ghost.io) explains that the voice model handles the conversation, while a separate "backend" model or agent does the thinking and work, such as looking up an order or using a tool[\[7\]](#cite-note-7). Sending a task to the backend is called delegation[\[7\]](#cite-note-7). The guide gives the example of a user who asks about an order and adds a detail while the backend checks its status; GPT-Live "can keep talking with the user and explain the result when it arrives"[\[7\]](#cite-note-7).

There are two ways to run the backend, according to OpenAI[\[7\]](#cite-note-7):

- **Responses delegation**, where OpenAI runs the backend model and passes context between it and the voice model.
- **Client delegation**, where a business connects its own model, agent or service.

The changelog shows GPT-Live-1 became generally available in the API on 10 September, priced at $0.05 per minute billed per second, with backend usage charged separately[\[2\]](#cite-note-2). Since then OpenAI has added nine more natural voices in German, French, Hindi, Japanese and Korean (16 October) and outbound calling over SIP, the standard used to connect internet calls to phone numbers, for all organisations (2 November)[\[2\]](#cite-note-2).

## Standby, explained

### What it does

Standby is a new setting for both GPT-Live-1 and GPT-Live-1.1[\[1\]](#cite-note-1)[\[2\]](#cite-note-2). It works only for sessions connected over WebRTC, a technology browsers and apps use for live audio and video[\[2\]](#cite-note-2)[\[3\]](#cite-note-3)[\[4\]](#cite-note-4).

OpenAI's [guide to managing GPT-Live sessions](https://developers.openai.com/api/docs/guides/live-conversations?ref=signal-engine.ghost.io) sets out how it works[\[4\]](#cite-note-4):

- When creating the session, the developer sets it to wait for speech before the model starts working.
- The developer sets an idle timeout, a whole number of seconds from 5 to 60\. If nobody speaks and the model says nothing for that long, the session goes into standby.
- During standby, "the connection and session ID stay the same". When the user speaks again, the session resumes and restores the conversation history.
- If no timeout is set, the model stays active after it first wakes until the session ends.

An optional setting controls how sure the system must be that someone is speaking before it wakes up. It runs from 0 to 1 and defaults to 0.5; higher values need stronger evidence of speech[\[4\]](#cite-note-4).

### The catches

The guide lists several practical limits[\[4\]](#cite-note-4)[\[5\]](#cite-note-5):

- **Muting is not standby.** "Muting the microphone leaves the session active," the guide says[\[4\]](#cite-note-4), so muted time is billed at the active rate[\[5\]](#cite-note-5).
- **The microphone must stay on.** The app must keep the microphone track active during standby[\[4\]](#cite-note-4). See also [what counts as active time](#what-counts-as-active).
- **A slower first reply.** "After standby, the first response to a short utterance may take longer," OpenAI says[\[4\]](#cite-note-4).
- **Long gaps.** For long gaps between conversations, OpenAI suggests closing the session and starting a new one when the user returns, rather than leaving it on standby[\[4\]](#cite-note-4).

## What it costs

### The rates

OpenAI's [pricing page](https://developers.openai.com/api/docs/pricing?ref=signal-engine.ghost.io) lists the same rates for both GPT-Live models[\[6\]](#cite-note-6):

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

Voice sessions are billed per second[\[6\]](#cite-note-6). The model page says the lower standby rate "applies to WebRTC sessions while inference is paused"[\[3\]](#cite-note-3). Backend model and tool usage is billed separately[\[3\]](#cite-note-3)[\[5\]](#cite-note-5).

### What counts as active time

OpenAI's [cost guide](https://developers.openai.com/api/docs/guides/voice-latency-cost?ref=signal-engine.ghost.io) says active time includes speech, "silence before the idle timeout expires", and time spent waiting for backend work while the model is running[\[5\]](#cite-note-5). So standby saves money only after the timeout has passed. A 30-second timeout means the first 30 seconds of every silence are still billed at the active rate.

The cost guide also describes a start-up charge. By default, creating a WebRTC session bills 15 seconds of voice time while the session starts, and that amount is credited against the session's charges once it is running[\[5\]](#cite-note-5). In OpenAI's example, a 90-second active session already includes those 15 seconds[\[5\]](#cite-note-5). The guide tells developers to account for this when deciding whether to close and reopen sessions[\[5\]](#cite-note-5).

### What that means in rands

These are our own calculations, using OpenAI's published rates[\[6\]](#cite-note-6) and the rand's close of R17.45 to the US dollar on Thursday 19 November reported by [Business Day](https://www.businessday.co.za/markets/2026-11-20-coalition-jitters-add-to-rand-and-jse-pain-as-global-pressure-mounts/?ref=signal-engine.ghost.io)[\[8\]](#cite-note-8). Exchange rates change daily, and your bank or card provider may charge a different rate.

| Time     | Active ($0.05/min)  | Standby ($0.02/min) |
| -------- | ------------------- | ------------------- |
| 1 minute | $0.05 (about R0.87) | $0.02 (about R0.35) |
| 1 hour   | $3.00 (about R52)   | $1.20 (about R21)   |
| 8 hours  | $24.00 (about R419) | $9.60 (about R168)  |

Here is an example, using the method in OpenAI's cost guide of multiplying active minutes by the active rate and standby minutes by the standby rate[\[5\]](#cite-note-5). Suppose a voice assistant on a shop's website is left connected for an eight-hour day, and people are actually talking to it, plus the silence before each timeout, for two of those hours:

- **Without standby:** eight hours at $3.00 an hour is $24.00, or about R419.
- **With standby:** two hours active ($6.00) plus six hours on standby ($7.20) is $13.20, or about R230.

That is a saving of about 45% on the voice charge in this example. Backend model and tool charges come on top in both cases[\[3\]](#cite-note-3)[\[5\]](#cite-note-5), and the real saving depends on how long the silences are and what timeout is set. For an assistant that is only opened when a customer calls and closed when they hang up, standby makes much less difference.

## Other details for developers

### Voices and languages

The sessions guide lists 21 additional voices with names, languages and accents[\[4\]](#cite-note-4). They include English voices with Australian, British, Irish, North American, Filipino and Southern US influences, plus Brazilian Portuguese, German, French, Hindi, Japanese and Korean voices[\[4\]](#cite-note-4). The default voice is called marin[\[4\]](#cite-note-4). The list does not include any South African language or a South African English accent. OpenAI advises developers to "test the voice with the languages and pronunciation your application needs"[\[4\]](#cite-note-4). Developers can also create custom voices from a text description, which the changelog says is free to create, with speech billed at normal GPT-Live rates[\[2\]](#cite-note-2).

### Instructions and limits

Instructions for how the assistant should behave can be up to 16,384 tokens long, according to the sessions guide[\[4\]](#cite-note-4). A token is a small chunk of text, often part of a word. The model, voice and delegation type are fixed when the session starts; changing them needs a new session[\[4\]](#cite-note-4). The model page says GPT-Live-1.1 and GPT-Live-1 have separate limits on how many sessions can run at once[\[3\]](#cite-note-3). It lists function calling (letting the model trigger actions in other software) as supported, but structured outputs and fine-tuning as not supported[\[3\]](#cite-note-3).

## What we could not verify

- **Quality claims.** OpenAI says GPT-Live-1.1 is more accurate and more consistent across languages[\[3\]](#cite-note-3), but it does not publish benchmark figures on the pages we read, and we found no independent tests.
- **The ChatGPT claim.** The statement that GPT-Live-1.1 has powered voice in ChatGPT Plus and Pro since 15 October appears only in the forum post[\[1\]](#cite-note-1).
- **Who posted the announcement.** We could not confirm whether the forum account "sps" belongs to an OpenAI employee[\[1\]](#cite-note-1).
- **Availability by country.** None of the pages we read say whether there are any country limits.
- **Endpoint list.** The model page lists many API endpoints, but our captured copy does not show clearly which ones support GPT-Live-1.1, so we have not repeated that list.

## Why it matters for your business (analysis)

Analysis This section is our analysis, based on OpenAI's published information above. It is not news reporting, and it is not financial advice.

### Voice assistants are getting cheaper to leave running

The headline price for a talking voice assistant has not changed: $0.05 a minute[\[2\]](#cite-note-2)[\[6\]](#cite-note-6). What has changed is the cost of silence (see [standby, explained](#standby-explained)). Before standby, a voice session that stayed open cost the same whether anyone was talking or not. Now, after a timeout, quiet time costs 60% less[\[6\]](#cite-note-6), on our arithmetic from the published rates.

That matters most for assistants that sit open for long periods with short bursts of talk, such as a help screen in a shop, a kiosk at reception or a phone line that stays open while a customer goes to find a reference number. For a business, the right question is not "what does a minute cost?" but "how much of a typical call is silence?"

### Work out your own numbers before you build

Before paying a developer to build a voice assistant, estimate:

- how many calls or conversations you expect in a day;
- how long a typical one lasts, and how much of it is silence;
- what the backend will cost for lookups and tools, which is billed separately[\[5\]](#cite-note-5);
- the exchange rate risk. All these prices are in US dollars, and the rand has weakened this month[\[8\]](#cite-note-8).

A simple spreadsheet with these four lines will tell you more than any headline rate.

### Test the accent and the language

The voice list has no South African language and no South African English accent[\[4\]](#cite-note-4) (see [voices and languages](#voices-and-languages)). That does not mean the model cannot understand South African callers. OpenAI describes it as a model for multilingual conversations, and more accurate transcription is one of the improvements it claims[\[3\]](#cite-note-3). But it does mean you should test it with real customers' voices, accents, place names and product names before relying on it. Pay particular attention to numbers, addresses and spellings, which the forum post says the new model handles better[\[1\]](#cite-note-1). Check that claim yourself.

### Keep a person in the loop

OpenAI's guide says your own application is still responsible for checking permissions, getting confirmations and running the functions that touch your systems[\[7\]](#cite-note-7). In plain terms, the voice assistant talks; your software decides what it is allowed to do. For a small business, the safest start is an assistant that answers questions and takes messages, with a person confirming bookings, refunds or payments.

### Mind the waiting time

Time spent waiting for backend work is billed at the active rate[\[5\]](#cite-note-5). A slow lookup costs money as well as patience. OpenAI's cost guide suggests giving the assistant the information it needs before the call starts, such as an order number and status, so the caller does not wait for another lookup[\[5\]](#cite-note-5). For a business, that means connecting the assistant to clean, up-to-date records is as important as choosing the model.

### What to do this week

- If you already use GPT-Live-1, ask your developer whether switching to `gpt-live-1.1` and turning on standby makes sense for your calls.
- If you are considering a voice assistant, write down how long your typical customer calls are and how much of each is silence.
- Record a few real customer questions in your customers' accents and languages, and use them to test any voice product before you buy.
- Budget in rands with a margin for exchange-rate moves.

## Sources

1. [↩a](#cite-ref-1-a) [↩b](#cite-ref-1-b) [↩c](#cite-ref-1-c) [↩d](#cite-ref-1-d) [↩e](#cite-ref-1-e) [↩f](#cite-ref-1-f) [↩g](#cite-ref-1-g) [↩h](#cite-ref-1-h) [↩i](#cite-ref-1-i) OpenAI Developer Community (post by "sps", Announcements), [“GPT-Live-1.1 has arrived in the API”](https://community.openai.com/t/gpt-live-1-1-has-arrived-in-the-api/1418537?ref=signal-engine.ghost.io), 20 November 2026.
2. [↩a](#cite-ref-2-a) [↩b](#cite-ref-2-b) [↩c](#cite-ref-2-c) [↩d](#cite-ref-2-d) [↩e](#cite-ref-2-e) [↩f](#cite-ref-2-f) [↩g](#cite-ref-2-g) [↩h](#cite-ref-2-h) [↩i](#cite-ref-2-i) [↩j](#cite-ref-2-j) [↩k](#cite-ref-2-k) OpenAI, [“Changelog”](https://developers.openai.com/api/docs/changelog?ref=signal-engine.ghost.io) (OpenAI API documentation), entries dated 10 September to 19 November 2026, read 20 November 2026.
3. [↩a](#cite-ref-3-a) [↩b](#cite-ref-3-b) [↩c](#cite-ref-3-c) [↩d](#cite-ref-3-d) [↩e](#cite-ref-3-e) [↩f](#cite-ref-3-f) [↩g](#cite-ref-3-g) [↩h](#cite-ref-3-h) [↩i](#cite-ref-3-i) [↩j](#cite-ref-3-j) [↩k](#cite-ref-3-k) [↩l](#cite-ref-3-l) [↩m](#cite-ref-3-m) [↩n](#cite-ref-3-n) [↩o](#cite-ref-3-o) [↩p](#cite-ref-3-p) OpenAI, [“GPT-Live-1.1”](https://developers.openai.com/api/docs/models/gpt-live-1.1?ref=signal-engine.ghost.io) (model page, OpenAI API documentation), read 20 November 2026.
4. [↩a](#cite-ref-4-a) [↩b](#cite-ref-4-b) [↩c](#cite-ref-4-c) [↩d](#cite-ref-4-d) [↩e](#cite-ref-4-e) [↩f](#cite-ref-4-f) [↩g](#cite-ref-4-g) [↩h](#cite-ref-4-h) [↩i](#cite-ref-4-i) [↩j](#cite-ref-4-j) [↩k](#cite-ref-4-k) [↩l](#cite-ref-4-l) [↩m](#cite-ref-4-m) [↩n](#cite-ref-4-n) [↩o](#cite-ref-4-o) [↩p](#cite-ref-4-p) [↩q](#cite-ref-4-q) OpenAI, [“Managing GPT-Live sessions”](https://developers.openai.com/api/docs/guides/live-conversations?ref=signal-engine.ghost.io) (OpenAI API documentation), read 20 November 2026.
5. [↩a](#cite-ref-5-a) [↩b](#cite-ref-5-b) [↩c](#cite-ref-5-c) [↩d](#cite-ref-5-d) [↩e](#cite-ref-5-e) [↩f](#cite-ref-5-f) [↩g](#cite-ref-5-g) [↩h](#cite-ref-5-h) [↩i](#cite-ref-5-i) [↩j](#cite-ref-5-j) [↩k](#cite-ref-5-k) [↩l](#cite-ref-5-l) OpenAI, [“Cost optimization”](https://developers.openai.com/api/docs/guides/voice-latency-cost?ref=signal-engine.ghost.io) (OpenAI API documentation, GPT-Live usage and costs section), read 20 November 2026.
6. [↩a](#cite-ref-6-a) [↩b](#cite-ref-6-b) [↩c](#cite-ref-6-c) [↩d](#cite-ref-6-d) [↩e](#cite-ref-6-e) [↩f](#cite-ref-6-f) [↩g](#cite-ref-6-g) [↩h](#cite-ref-6-h) OpenAI, [“Pricing”](https://developers.openai.com/api/docs/pricing?ref=signal-engine.ghost.io) (OpenAI API documentation, GPT-Live section), read 20 November 2026.
7. [↩a](#cite-ref-7-a) [↩b](#cite-ref-7-b) [↩c](#cite-ref-7-c) [↩d](#cite-ref-7-d) [↩e](#cite-ref-7-e) OpenAI, [“GPT-Live”](https://developers.openai.com/api/docs/guides/live?ref=signal-engine.ghost.io) (OpenAI API documentation guide), read 20 November 2026.
8. [↩a](#cite-ref-8-a) [↩b](#cite-ref-8-b) Business Day (Andrew Linder), [“Coalition jitters add to rand and JSE pain as global pressure mounts”](https://www.businessday.co.za/markets/2026-11-20-coalition-jitters-add-to-rand-and-jse-pain-as-global-pressure-mounts/?ref=signal-engine.ghost.io), 20 November 2026.

This article is for information only and is not financial or investment advice.

Spotted an error? Tell us and we'll correct it.