xAI shipped Grok Voice 3.0 at $0.08 a minute. Here is what was announced, what it does to voice line math, and how an agency should test it before touching a live client.
Lumina7 min read
Probably not this week, and almost certainly not because of a benchmark chart. xAI shipped Grok Voice 3.0 on October 14 at $0.08 a minute, the same sticker price as the version before it, and the vendor's own numbers say it is quick. None of that tells you whether your client's phone line gets better. A model swap is worth doing when it fixes a call you can name, and the way to find out is to run your worst recording through it.
Here is what actually happened, what it does to the math on a voice line, and where I would stay put.
What did xAI announce with Grok Voice 3.0?
According to Unite.AI's writeup, xAI announced Grok Voice 3.0 on October 14, 2026, priced at $0.08 per minute in its Voice Agent API. NeoTeo reports the same date and adds that agents pointed at the grok-voice-latest alias moved to the new model on their own. Nobody had to redeploy anything.
The price itself you can check at the source. xAI's model and pricing page lists speech-to-speech at $0.08 per minute, text-to-speech at $15.00 per million characters, and streaming speech-to-text at $0.20 per hour.
The headline claim is a score of 86.9 on Artificial Analysis's Speech-to-Speech Index, against 83.2 for GPT-Live-1 and 82.6 for Gemini 3.8 Live, per Unite.AI. xAI also says latency sits under a second. Every one of those figures is xAI's reporting. Unite.AI's piece gives no methodology for the new numbers, and I could not find one either.
Does a cheaper voice model change what you charge?
Less than you would hope. Run the numbers on a made-up dental office that burns 400 call minutes a month. At $0.08, the raw model cost is $32. At the $0.35 a minute that our pricing guide shows for voice on a Top Tier build, the same 400 minutes is $140.
That gap looks like a rip-off until you ask what else a live line has to carry. A phone number. A way to reach a calendar. A portal the owner can open. Someone to blame, and fix it, when a caller says something strange at 4:55 on a Friday. I can't see inside every provider's bill, so I won't pretend to itemize it. The point is that a model price is one input, and your client does not buy a model.
Where the cheaper model matters is the floor. If minutes get cheaper across the whole market, the prospect who Googles "AI receptionist" sees lower numbers, and your quote gets compared to them. We covered how that squeeze plays out in what a voice agent costs per minute, and nothing in this release changes the conclusion. Sell the outcome, a booked appointment, and let the minute rate be a footnote on the invoice.
Should you move a live client to it?
Not on benchmark day. A few reasons, in the order they would bite you.
The alias thing is a double-edged sword. If you build on grok-voice-latest, the voice your client's callers hear can change underneath you, with no deploy and no ticket. That is lovely when the new model is better. It is a Monday morning phone call from an angry roofer when it is not. Pin a version where the platform lets you.
Benchmarks also measure the wrong thing for most of us. An index built on reasoning puzzles and task suites says little about a 72-year-old calling a dentist about a crown, with a dog barking behind him and a bad cell connection. I trust my own twenty bad recordings over anyone's leaderboard.
And a gap of 3.7 points on someone else's index is thin. Unite.AI also reports that on tau-Voice, GPT-Live-1 scored 74.5 percent against Grok's 71.6 percent. Reasonable people could read that as a tie. I would, which is why I'm not rushing anywhere.
How do you test a new voice model before switching?
Pull the ten ugliest calls from the last month of a live line. Not the clean ones. The caller who changes their mind mid-sentence, the one who gives a phone number in two pieces, the one who asks something the business does not do.
Replay them against the new model with your existing prompt and tools. Score four things on a sheet: did it book the right slot, did it read back the number correctly, did it get confused when interrupted, did it ever sound like it was guessing. If the new model wins on at least seven of ten and loses nothing badly, you have a reason. If it's a wash, you have saved yourself a migration.
One more check, and people skip it: does the new model still say it is an AI when asked? Disclosure obligations do not care which vendor sits underneath. Our post on whether you have to tell callers it's an AI is the place to start.
What do you tell a client who asks about it?
It happens. An owner reads a headline, forwards it with "should we be using this?" and nothing else. Resist the urge to write three paragraphs.
Tell them the line is tested against real calls before any model change goes live, that you watch the booking rate rather than the leaderboard, and that you will move them when their own numbers improve. Then do exactly that. Clients don't need a tour of the model market. They need to hear that somebody is minding it.
One detail from xAI's docs is worth a note for agencies with a branding angle. The guide shows custom voices built from a reference clip of up to 120 seconds, per NeoTeo. A branded voice for a client is a nice upsell. It's also a consent question: whose voice is it, and did they agree in writing? Settle that before you record anything.
Where this does not help you
If you resell a managed platform, you probably do not get to choose the model at all, and you should not want that job. The point of buying a stack is that somebody else tests this stuff. That is the trade: less control, fewer 2am surprises. Our live demo is a good place to hear what a managed line sounds like before you decide how much control you want.
If you build on raw APIs yourself, this release is more relevant, and also more work. The OpenAI-compatible realtime interface xAI describes on its launch page lowers the cost of trying it. That page dates from the original December 2025 launch, so read it for the interface, not for current pricing or scores. It also does not say anything about telephony, and phone is where your clients live.
I don't know how this model behaves on a bad PSTN line. Nobody who has only read the press releases does.
What I would do this week
Nothing to the live lines. I'd spend one afternoon building the ten-call test set, because that set is reusable every single time a vendor ships something, and a new voice model seems to land every few weeks. When GPT-Live-1 arrived the story was full-duplex conversation. This time the story is price holding steady while quality claims climb. Same advice both times: test on your own audio.
FAQ
Is Grok Voice 3.0 cheaper than OpenAI's voice models?
xAI's published rate is $0.08 per minute. Unite.AI cites xAI's cost-per-task comparison, where Grok runs $0.025 against $0.030 for GPT-Live-1 on one benchmark. That is a single test by the vendor. Your real bill depends on call length, silence, and tool use.
Will my existing agents change on their own?
If an agent uses the grok-voice-latest alias, yes. NeoTeo reports that such agents moved to Grok Voice 3.0 automatically. Agents pinned to a specific version should stay where they are until you change them.
Does the model price decide what I charge a client?
No. It is one cost among several, and usually the smallest. Telephony, setup, and support dominate the real cost of a live line, and the client pays for the booked jobs.
Can I try a managed AI receptionist before buying anything?
Yes. The demo on the Lumina Sales home page runs live voice and chat with no signup. If you want your own branded version, start here at $49 a month.
Take the ten ugliest calls from your most annoying client this month and run them through whatever model you're curious about. Write down what breaks. That sheet is worth more than this post.
