Somewhere in a Starlink call centre, a customer is explaining that their dish is pointing at a tree. Nobody is going to roll their eyes at this. Nobody is going to sigh into a headset. The call is being handled by an AI agent, and according to the people who built it, a lot of these calls end without a human ever joining.

On October 14, xAI released Grok Voice 3.0, its latest speech-to-speech model for voice agents, and published numbers from its own deployments. Vendor numbers deserve a raised eyebrow, always. But they are also a useful look at where the bar is moving, and the bar is now sitting at your front desk.

What xAI Actually Announced

Grok Voice 3.0 is available now through xAI's Voice Agent API at $0.08 per minute, which works out to $4.80 an hour. That is the same price as Grok Voice 2.0, so the upgrade costs nothing extra. Agents already pointed at the latest-voice alias were switched over automatically.

xAI says the model responds in under a second and handles pauses, interruptions and turn-taking, scoring 97.0% on the Full Duplex Bench. On Artificial Analysis's Speech-to-Speech Index it ranks first with 86.9, ahead of OpenAI's GPT-Live-1 at 83.2 and Google's Gemini 3.8 Live at 82.6, according to the table xAI published.

It is not a clean sweep, and to xAI's credit the page shows it. On τ-Voice, a test built around support calls with accents, background noise and interruptions, Grok Voice 3.0 scored 71.6% against 74.5% for GPT-Live-1. xAI describes the two as comparable, with Grok about 40% lower on latency.

The Starlink Numbers

Here is the part worth reading twice. xAI says Starlink's call centres take 20,000 customer calls a day across multiple languages, and about 70% are resolved with no human in the loop. One in four sales calls ends in a purchase before the caller hangs up.

One agent, per xAI, verifies the customer, troubleshoots the hardware, issues credits and ships replacement equipment. That is not an agent that says "let me take a message." That is an agent that does the job.

Tesla is the other named customer. xAI says the model handles millions of conversations a day hands-free in Tesla vehicles, from checking a calendar to booking dinner. Cute, but not your problem. The Starlink example is the one that maps onto a business with a phone that rings.

Read the Fine Print Like a Grown-Up

These are xAI's figures about xAI's customers, published on xAI's own site. Starlink is a huge operation with structured problems: verify an account, check a dish, send a replacement. A plumber taking an emergency call at 6pm on a Friday is a different animal, and the benchmarks do not tell you how an agent copes with it.

Also, $4.80 an hour is the price of the model, not of a working phone system. Someone still has to write the instructions, connect the booking calendar, set the handoff rules and listen to the calls afterwards. The model is the engine. You still need the rest of the car.

ServiceNow, whose open-source EVA-Bench tests voice agents on enterprise scenarios, was quoted on the release as saying the result sets "a new standard for what enterprises can expect from AI on the front line of customer service." Notice the word "expect." Expectations travel.

Why This Lands on a Small Business

A year or two ago, an agent that could interrupt gracefully, handle accents and actually finish a task was a research demo. Now it is a line item on an API price list, and the same rate applies whether the caller is a Starlink customer or the person ringing your dental practice about a toothache.

The effect is not that every small business gets Starlink's results next month. It is that your customers will quietly adjust what a phone call should feel like. Answered immediately, no hold music, no "please leave a message." The first time a caller meets that somewhere else, your voicemail starts to feel like a landline in a museum.

So the practical question is not which model is number one this week. That will change by the time you finish your coffee. The question is how many calls you currently miss, what each one is worth, and which of those calls are simple enough to hand off, such as booking, hours, quotes and basic questions. Count those first. Keep a human for the delicate ones. Then pick the tooling that gets the simple calls answered, and judge it on your own call recordings, not on anybody's leaderboard.