In this guide · 8 sections
- 01What OpenAI released on 15 and 16 October 2026
- 02What prompt-to-voice means for a business
- 03What a voice agent costs per minute
- 04What you need to connect it to a phone line
- 05Singapore rules: disclosure, recording and marketing calls
- 06Is a voice agent worth it for a small business?
- 07Do it yourself or get help?
- 08Frequently asked questions
Short answer
Yes, technically. OpenAI's GPT-Live 1 costs US$0.05 per minute of voice session (billed per second), plus backend model and telephony charges, and can take phone calls through SIP or a partner like Twilio. Since 16 Oct 2026 you can design a voice from a text prompt. Disclose the AI and follow DNC rules.
What OpenAI released on 15 and 16 October 2026
Two separate updates landed in the same week. OpenAI's API changelog lists the developer changes on 16 October, and the ChatGPT release notes cover the consumer app on 15 October.
| Date | Where | What changed |
|---|---|---|
| 10 Sep 2026 | API | GPT-Live 1 generally available: full-duplex voice conversations, US$0.05 per minute |
| 15 Oct 2026 | ChatGPT | GPT-Live-1.1 and a mini version in ChatGPT Voice; new voices in several languages |
| 15 Oct 2026 | API | Native Decisions integration in GPT-Live, so an app can act while the caller is speaking |
| 16 Oct 2026 | API | Prompt-to-voice for all API customers: create a custom voice from a text description |
| 16 Oct 2026 | API | Nine new GPT-Live voices in German, French, Hindi, Japanese and Korean |
For businesses, the API changes are the ones that matter. ChatGPT Voice is a personal assistant; it does not answer your company phone. A phone agent is something you build on the API, or buy from a platform that builds on it.
What prompt-to-voice means for a business
OpenAI's custom voices guide says prompt-to-voice creates and saves a voice from a text description, such as "a calm, friendly receptionist with a neutral accent". You do not need a reference recording or a consent recording for this route. Creating the voice is free; speech uses normal GPT-Live session billing, and an organisation can save up to 100 prompt-generated voices.
Cloning a real person's voice from a recording is different. The same guide says recording-based custom voices are limited to eligible customers through OpenAI's sales team and need a consent recording in which the voice owner reads a fixed consent phrase.
What a voice agent costs per minute
OpenAI's API pricing page lists GPT-Live 1 at US$0.05 per minute, billed per second without rounding up. The backend model that looks up answers or calls tools is charged separately, at its normal token rates. If you build on the older Realtime models instead, they are billed per token of audio.
| Item | Price (USD, Oct 2026) | Notes |
|---|---|---|
| GPT-Live 1 voice session | US$0.05 per minute | Billed per second |
| Backend model (answers, tools) | Normal token rates | Depends on model and how much it reads |
| gpt-realtime-2.1 audio | US$32 input / US$64 output per 1M tokens | Alternative speech-to-speech model |
| gpt-realtime-2.1-mini audio | US$10 input / US$20 output per 1M tokens | Cheaper alternative |
| Phone number and call minutes | Set by your telephony provider | Check your SIP or Twilio-type provider |
If you would rather not build anything, our guide to AI voice receptionists in Singapore compares ready-made platforms, their per-minute prices and how well they handle local accents.
What you need to connect it to a phone line
OpenAI's telephony and SIP guide describes two ways in: a direct SIP connection, where your phone provider sends call audio straight to OpenAI, or an application that relays audio between the call and the model. OpenAI also lists partner integrations for LiveKit, Twilio, Telnyx, Bandwidth and Daily/Pipecat.
Get a phone number that supports SIP
Use a provider that offers SIP trunking or a programmable voice API. Many businesses forward their existing number to it after hours or when the line is busy.
Enable GPT-Live SIP for your OpenAI project
OpenAI's guide says to confirm SIP support is enabled for the project and that the provider's trunk routes to it. SIP signalling uses TLS and audio must use SRTP.
Build the backend that decides what happens
Your server accepts or rejects each call, sets the instructions and voice, and handles tools such as checking a booking calendar. This is where most of the work is.
Write a narrow script
Limit the agent to a few jobs: opening hours, prices from your list, taking a message, booking or rescheduling. Tell it to hand over to a human for anything else.
Plan the handover
Decide how a caller reaches a person: transfer to a mobile, a callback promise, or a WhatsApp message to the owner.
Test with real calls before going live
Have staff call with typical questions, background noise and Singlish phrasing. Check what it says when it does not know.
Singapore rules: disclosure, recording and marketing calls
There is no special Singapore licence for AI phone agents, but existing rules still apply. This is general information, not legal advice.
Before your voice agent takes real calls
An AI voice making promotional calls is treated like any other marketing call. You must check Singapore numbers against the DNC Registry first, keep consent records and offer an opt-out. Our DNC Registry check guide explains the process. Reminders about an existing appointment are service calls, not marketing.
Is a voice agent worth it for a small business?
It depends on how many calls you miss and how predictable they are. If most calls ask the same five questions, an AI agent can handle them well. If every call is different and needs judgement, a good voicemail with a fast callback may be cheaper and less risky.
Worth it if you have
- Many repetitive calls: hours, location, prices, bookings
- Missed calls after hours or during busy periods
- Callers who speak languages the model handles well (test yours)
- A booking system the agent can check and update
Probably not yet if
- Few calls a day; a person can answer them
- Calls need judgement, negotiation or sensitive information
- No one has time to review transcripts and fix the script
- Most enquiries already arrive by WhatsApp or web chat
Many Singapore customers message instead of calling. If that is true for you, a text assistant on your website or WhatsApp is cheaper to run and easier to check. Our comparison of ChatGPT, Claude and Gemini for business helps if you are still choosing which AI vendor to build on.
Do it yourself or get help?
A developer comfortable with webhooks and SIP can build a basic voice agent from OpenAI's quickstart in a few days. The harder parts are a reliable handover to humans, connecting your booking system, and testing enough real calls to trust it.
If text chat suits your customers better, our AI Website Assistant (S$790) answers questions from your own FAQ, price list and policies on your website. IT Sorted is a Singapore-based team of developers with over 10 years of experience. You pay upfront, send your brief in writing through the site chat or contact form, and follow progress on your order page.
Frequently asked questions
OpenAI's API changelog lists two voice updates on 16 October 2026: prompt-to-voice for all API customers, which creates a custom voice from a text description, and nine new natural voices for GPT-Live in German, French, Hindi, Japanese and Korean. Creating a prompt-generated voice is free, and speech uses normal GPT-Live session billing.
OpenAI's pricing page lists GPT-Live 1 at US$0.05 per minute, billed per second, as of October 2026. Backend model and tool usage are charged separately at token rates, and your telephony provider charges for the phone number and call minutes. A three-minute call therefore costs about US$0.15 in voice session time plus those extras.
Yes, with some setup. OpenAI's telephony guide describes connecting phone calls through a SIP trunk that sends call audio to OpenAI, or through an application that relays audio. It lists partner integrations including Twilio, Telnyx, LiveKit, Bandwidth and Daily. You still need a backend server to accept calls, set instructions and handle bookings or transfers.
OpenAI's custom voices guide tells developers to let listeners know that speech is AI-generated and not to impersonate anyone. It is also good practice for customer trust. Singapore's PDPA separately requires you to tell callers if you record or transcribe calls and why. A short greeting such as "You are speaking with our AI assistant" covers the disclosure.
Only within the Do Not Call rules. An AI voice making promotional calls is treated like any other marketing call, so you must check Singapore numbers against the DNC Registry before calling unless the person has given clear consent, keep records and provide an opt-out. Reminder calls about an existing appointment are service calls and are not covered by DNC.