Tech

Grok Voice 3.0 wants your support calls. Read the benchmarks first.

xAI calls its new voice model the world’s most intelligent. On simulated support calls, a rival still scored higher, and its showcase customers are Musk companies.

Published

4 hours ago

October 21, 2026

Massih Shahbazi

A headset resting on a laptop keyboard, in black and white, beside the headline Nobody picks up

The next time you call Starlink support, there’s a decent chance nobody human picks up. xAI says that’s a feature.

xAI released Grok Voice 3.0 on Oct. 14, a speech-to-speech model for building voice agents through its Voice Agent API, Unite.AI reported. It costs $0.08 a minute, or $4.80 an hour, the same price as Grok Voice 2.0, according to the same report. xAI calls it “the world’s most intelligent voice model.”

I went through the benchmarks xAI published alongside the launch, because that claim deserves a closer look than the press release gives it.

What it is

A speech-to-speech model listens to audio and answers in audio directly, rather than turning speech into text, thinking in text and reading the result back. Chaining those steps together is part of what made older voice assistants feel slow and stilted. The goal is a conversation that sounds natural: fast replies, handling interruptions, knowing when someone has finished talking.

xAI says Grok Voice 3.0 responds in under a second, according to Unite.AI, though the coverage doesn’t include an independent latency test. Developers using the grok-voice-latest alias were switched to the new model automatically, and xAI deprecated 2.0, a developer post on DEV Community noted. That’s convenient if the new model is better, and a headache if a voice agent that worked fine on Monday suddenly behaves differently on Wednesday.

The scoreboard

On the Artificial Analysis Speech-to-Speech Index, cited by xAI, Grok Voice 3.0 scored 86.9, ahead of GPT-Live-1 at 83.2, Gemini 3.8 Live at 82.6 and its own predecessor at 81.3, according to Unite.AI.

That’s the number behind the “most intelligent” claim. But it isn’t the whole picture.

On tau-Voice, a benchmark that simulates customer-support calls, Grok Voice 3.0 scored 71.6%, Unite.AI reported. GPT-Live-1 scored 74.5%. Gemini 3.8 Live scored 68.6%, and Grok Voice 2.0 scored 56.5%.

So on the test closest to what xAI is selling, handling support calls, a rival model came out ahead. xAI described 3.0 as “comparable” to GPT-Live-1 with 40% lower latency, according to Unite.AI. That’s a fair framing for a three-point gap if speed matters more to you than accuracy. It isn’t the same as being best.

The jump from 2.0 is real, though. Going from 56.5% to 71.6% on simulated support calls is a big improvement in a single version.

xAI also reported a top accuracy score of 74% on ServiceNow’s EVA-Bench, and ServiceNow’s Nitin Aggarwal said the model “sets a new Pareto frontier on EVA-Bench,” according to Unite.AI.

The real product: call centers

The deployment claims tell you where the money is. xAI says Starlink’s call centers handle about 20,000 customer calls a day across multiple languages with Grok Voice, that roughly 70% are resolved without a human, and that one in four sales calls ends in a purchase before the caller hangs up, according to Unite.AI. xAI also says the model handles millions of conversations a day in Tesla vehicles.

All of those figures come from xAI. None has been independently verified in the coverage I read. And look at who the customers are: Starlink and Tesla are Elon Musk’s companies, and so is xAI. The flagship case studies for this product are its sister companies.

From the caller’s phone

For a customer, a voice agent that resolves 70% of calls sounds great when you’re in the 70%. The question is the other 30%. If an agent can verify your account, troubleshoot hardware, issue credits and ship replacement equipment, as xAI says one Starlink agent does, it can also deny those things. How easy it is to reach a person when the machine is wrong is the part no benchmark measures.

For workers, the math is blunter. At $4.80 an hour, a voice model costs a fraction of a human support agent’s wage. Companies will run that calculation whether or not the model is the smartest on the market. Good enough and cheap is a business case on its own.

What we don’t know

Most of the performance claims rest on benchmarks reported through xAI’s own announcement. Some lower-quality blogs have repeated additional figures that couldn’t be traced to a primary source. The coverage doesn’t list which languages the model supports, and xAI’s latency and deployment numbers haven’t been tested by anyone outside the company, as far as the reporting shows.

What’s clear is the direction. The voice in your car, and soon the voice on the support line, is being tuned to sound human, answer fast and close the ticket. Whether it actually helps you is a separate question, and you’ll only find out when you call.