---
title: "The Voice on the Other End of the Line"
paper: "The Daily Model"
edition: 2026-10-26
newsroom: xAI
edition_id: 2026-10-27T05-55-05-730Z-906f805df7-approved
content_hash: 77a22d0e5868f396d31ed49096366b28975011d51c5777716d5d4e2dc9a98bb8
sources:
  - Introducing Grok Voice 3.0 (https://x.ai/news/grok-voice-3) 2026-10-22
---

# The Voice on the Other End of the Line

The phone rings at a Starlink call center, and a voice answers. It is not a person. According to xAI, this happens 20,000 times a day.

That is the deployment claim behind Grok Voice 3.0, announced October 22. xAI says Starlink call centers use Grok Voice to answer 20,000 customer calls daily across multiple languages, with roughly 70% resolved with no human in the loop. One in four sales calls, the company says, ends in a purchase before the caller hangs up. A single agent built on Grok Voice 3.0 can verify customers, troubleshoot hardware, issue credits, and ship replacement equipment—though xAI does not say every call involves all of those steps, or that every one of the 20,000 runs on the new version. The company also says Grok Voice handles millions of conversations a day in Tesla vehicles.

These are vendor numbers, and they deserve that label. But the shape of the claim matters more than the polish. A support line is not a demo. It is accents, background noise, interruptions, and people who did not call because they were happy. xAI's own benchmark table concedes the difficulty: on τ-Voice, which measures exactly those conditions, Grok Voice 3.0 scores 71.6%—described as comparable with GPT-Live-1 (Astra), though xAI says at 40% lower latency.

The headline figures come from Artificial Analysis, which xAI describes as independent testing. Grok Voice 3.0 ranks first on the Speech-to-Speech Index at 86.9, ahead of GPT-Live-1 Astra at 83.2 and Gemini 3.8 Live at 82.6. It scores 97% on Full Duplex Bench, the turn-taking measure, and 94.8% on task success. On ServiceNow's EVA-Bench for enterprise voice agents, xAI claims the top accuracy score at 74%, with a quoted endorsement from ServiceNow's agentic evals lead. Cost per task on Big Bench Audio: $0.025, against $0.030 for GPT-Live-1 and $0.055 for GPT-Realtime-2.1.

For builders, the practical facts are simpler. The API price holds at $0.08 per minute, the same as Grok Voice 2.0, and agents configured to use grok-voice-latest were switched automatically—no code changes. Latency is claimed under one second. If your agent uses grok-voice-latest, you are running this model now, whether you benchmarked it or not.

That last point cuts both ways. Automatic upgrades are convenient until a regression appears in a call flow you did not design the model for. And the 70% resolution figure leaves 30% of calls going somewhere—xAI does not say where, or what a failed call costs the person on the other end.

The voice on the line is, on xAI's numbers, more capable than its predecessor at the same price. The interesting question is no longer whether it can talk. It is what happens on the calls where talking is not enough.

[Official source](https://x.ai/news/grok-voice-3)