Front Desk now answers on Grok Voice 3. Here's what that means for your callers

Most calls to a contractor or an injury firm don't come from a quiet room. A homeowner calls about a dead air conditioner with the kids and the television (TV) going in the background. A crew lead calls from a job site. Someone calls an injury firm two days after a crash, from a kitchen where the news is on. Whatever answers that call has to hear the person, keep track of what they said, and get the details right.

That's the bar we hold Front Desk, our artificial intelligence (AI) phone receptionist, to. We've been testing Grok Voice 3, the speech-to-speech model xAI released on October 14, 2026, on Front Desk calls, and we've now moved over to it.

What Grok Voice 3 is

Grok Voice 3 is a speech-to-speech model. It listens to the caller and answers in voice directly, rather than turning speech into text, writing a reply, and reading it back. xAI prices it at $0.08 per minute, the same as the previous version, according to its launch post.

Here's how the outside testers scored it. These are their numbers, not ours.

• Artificial Analysis ranks it first on its Speech to Speech Index at 86.9, ahead of OpenAI's GPT-Live-1 at 83.2. It's second on Tau Voice, a test of support calls with accents, background noise, and interruptions, and its first audio comes back in 0.81 seconds. (Artificial Analysis)
• Cekura, which runs scripted test calls against voice agents, reports it passed 81.7% of 82 scenarios, compared with 79.3% for OpenAI's GPT Realtime 2.1. (Cekura)
• Veris, in its VAmoS Pro benchmark, had Grok Voice at the top for task completion at 44.7%. With a TV playing in the caller's room it still completed 33.8%. Across the 14 setups Veris tested, average completion dropped from 38.7% to 8.6% with the TV on. Veris notes its samples are small and that Grok Voice and GPT-Live-1 are statistically tied at the top. (Veris)

Why background noise is the test that matters

The Veris result is the one we keep coming back to. A TV in the room is not an edge case. It's a normal Tuesday evening. When an agent mistakes a voice on the TV for the caller, it stops talking mid-sentence, loses its place, or answers a question nobody asked. The caller hangs up, and the next shop or firm on their list gets the job.

A model that holds up when the room is loud means fewer of those calls fall apart halfway through.

What it means for a home services caller

For a heating, ventilation, and air conditioning (HVAC) shop, a plumbing shop, or an electrical shop, a good call ends with a job on the schedule. Front Desk answers the business line 24/7, asks what's wrong, where, and how urgent it is, takes the booking details and confirms them with your team, and flags emergencies so whoever is on call can call back fast. Every call is recorded and transcribed.

A stronger voice model helps in the places where calls usually go wrong: hearing the address over a barking dog, reading a phone number back in the right order, and staying on track when the caller interrupts to add "oh, and it's also leaking."

What it means for a personal injury caller

For a personal injury firm, the first call decides whether there's a case at all. The caller is often hurt, stressed, and not in a quiet place. Front Desk is built to answer around the clock, gather the accident details, injuries, and contact information, and flag serious injuries so someone at the firm calls back quickly. It tells callers it's an AI and gives no legal advice, and an attorney reviews every intake.

Getting a name, a date, and a callback number right on the first try matters more here than almost anywhere. A better listener on a bad line is a direct help.

What we're not claiming

Benchmarks are tests, not your phone line. The numbers above are from Artificial Analysis, Cekura, and Veris, and they measure their own scenarios. We aren't publishing our own call statistics in this post. What we can say is that we tested it on Front Desk calls ourselves before moving over, and it's the model that did best on the outside tests closest to the calls our customers actually get.

See how it sounds

If you run a shop or an injury firm and want to hear how Front Desk handles a call, take a look at front-desk.ai.