xAI (X) releases Grok Voice 3.0 for building phone and web voice agents
1 · What happened
A new Grok voice model for real-time voice agents that can answer, talk and call your tools; xAI lists it at $0.08 per minute. This is a launch, not a customer result.
These numbers come from the builder and haven't been checked independently.
xAI · Oct 2026 · Source: docs.x.ai ↗
2 · How it works, in plain words
On 14 October 2026 xAI, the company behind Grok on X, made grok-voice-3.0 available in its Speech to Speech (voice agent) API at the same price as Grok Voice 2.0. The grok-voice-latest alias now points to it. xAI lists the API at $0.08 per minute, with tool calling, web and X search, and demo apps for web and Twilio phone agents.
3 · Setup steps
How to do it yourself
Our suggested steps. The company didn't publish theirs.
- 1
Create an xAI API account and API key at console.x.ai.
- 2
Start from xAI's Twilio phone agent or web agent demo app linked in the Voice docs.
- 3
Set the model to grok-voice-latest (now Grok Voice 3.0) and paste the prompt below as the agent's instructions.
- 4
Add one simple tool first, for example 'check opening hours' or 'take a message'.
- 5
Call it yourself 10 times with real questions and fix the instructions before giving the number to customers.
4 · The prompt
The instructions to give the AI
Paste this into Grok (xAI), swap the [brackets] for your details, and test it on a few real examples.
This is one part of the setup. It won't connect to your phone, inbox or other apps on its own; the steps above cover that.
Agent instructions (prompt)
You are the phone receptionist for [business name], a [type of business] in [city]. GOAL Answer callers politely, help with common questions, and take a message when you can't help. WHAT YOU KNOW - Opening hours: [hours] - Address: [address] - Services and prices: [list] - Common questions and answers: [paste 5-10 real Q&As] HOW TO TALK - Short, friendly sentences. One question at a time. - Speak the caller's language if they switch. - Never guess prices, dates or availability you were not given. WHEN YOU CAN'T HELP - Say so honestly, then take the caller's name, phone number and reason for calling, and read it back to confirm. NEVER - Promise refunds, bookings or anything not listed above. - Ask for card numbers or passwords.
[Highlighted] = replace with your own details before you run it.
What you should get back: A voice agent that answers calls or web voice chats in your business's name, answers common questions and takes messages.
Written by Quickly and Good, inspired by this case.
More detail: tools, prices and what you'll need
The company used: Grok Voice (xAI)
You could use: xAI API key (Grok Voice, grok-voice-latest), Twilio or a web page for calls, optional LiveKit
- Grok (xAI)
xAI's AI model, also used in X; its voice API builds real-time voice agents.
Voice agents $0.08/min (xAI API) · Official site ↗
- Twilio
Phone numbers and calling/SMS APIs used to connect voice agents to a real phone line.
Pay as you go · Official site ↗
Prices from each tool's own pricing page, checked Oct 2026.
What you'll need
- ☐An xAI API key (paid per minute)
- ☐A phone number via Twilio, or a web page
- ☐Basic coding or a developer's help
- ☐A short list of your real customer questions
Can I do this myself?
Possible if you can run a demo app from GitHub; otherwise ask a developer for a day.
Similar cases
Kogan rebuilt its customer-service bot on Agentforce in six weeks
Kogan.com moved its service bot to Agentforce, reaching production in about six weeks with 10 use cases including order tracking, recommendations, returns and identity checks.
Read how it worksCMA CGM's AI voice agents resolve 46% of cargo release calls
Shipping group CMA CGM uses HappyRobot voice agents to answer calls about cargo release, customs and documents, and to call customers about new routes and unpaid invoices. It is now scaling them up after pilots.
Read how it worksElevenLabs' own AI support agent resolves 65% of its tickets
ElevenLabs runs its customer support on its own agent platform. The agent reads text, screenshots and PDFs, looks up accounts, takes actions and hands hard cases to people. It resolves over 22,000 tickets a month.
Read how it works