Grok Voice 3.0 is xAI’s new speech-to-speech model, and it costs the same $0.08 per minute as the version before it. xAI released it on October 14, 2026, through its Voice Agent API. If you build voice apps, you may already be using it without knowing.
The short version: it’s faster, handles interruptions better, and didn’t raise the price. Most of the performance numbers come from xAI itself, so treat them as claims until you’ve tested your own use case.
- Released: October 14, 2026, in the xAI Voice Agent API
- Price: $0.08 per minute, or $4.80 per hour, unchanged from 2.0
- Upgrade: automatic if your agent uses the
grok-voice-latestalias - Already in use: Tesla voice features and Starlink support calls
Those four points cover most of what developers need to know. The rest of this post explains what changed and what to check.
What is Grok Voice 3.0?
It’s a speech-to-speech model. That means it listens to audio and replies with audio directly, instead of turning speech into text, thinking, and then reading text back. Fewer steps usually means less delay.
According to Blockchain.News, xAI’s release notes call it the current flagship speech-to-speech model. It streams audio and text both ways over WebSocket. Browser and app clients can sign in with short-lived tokens, so you don’t have to expose an API key.
Grok Voice 2.0 is now deprecated. The endpoint, session settings and events stay the same, so the move shouldn’t need code changes.
How much better is it than 2.0?
Here are the numbers xAI cites, as reported by Technobezz. Some were tested with Artificial Analysis, but most are still company-reported.
| Test | Grok Voice 2.0 | Grok Voice 3.0 |
|---|---|---|
| Speech-to-Speech Index | 81.3 | 86.9 |
| Full Duplex Bench (interruptions) | 95.1% | 97% |
| Price per minute | $0.08 | $0.08 |
The Full Duplex score matters most for real calls. It measures how well the model copes when you interrupt it or pause mid-sentence. Voice bots that talk over you are a big reason people hang up.
xAI also says responses arrive in under one second and that latency is 40% lower than OpenAI’s GPT-Live-1. It’s worth timing that yourself on your own network.
Where Grok Voice 3.0 still trails the competition
It doesn’t win everything. On τ-Voice, a support-call test, Grok Voice 3.0 scored 71.6%. GPT-Live-1 (Astra, Medium) scored 74.5%, according to the same Technobezz report.
Cost is where xAI looks better. On the Big Bench Audio task, the average cost was $0.025, compared with $0.030 for GPT-Live-1 and $0.055 for GPT-Realtime-2.1 (High). Cheaper and slightly less accurate on support calls is a trade many teams will accept, but not all.
If you compare voice models for pricing, our guide to ChatGPT audio uploads and limits shows how OpenAI handles audio on the consumer side.
Who is already using it?
xAI says the model runs in Tesla vehicles, where it handles millions of conversations a day, and in the Grok app’s voice mode. It also answers about 20,000 Starlink customer calls daily.
The standout claim is that around 70% of those Starlink calls end without a human agent. That figure comes from xAI, not an outside audit. Still, a live deployment at that scale tells you more than a lab score does.
ServiceNow’s EVA-Bench also ranked it first, with 74% accuracy, for both accuracy and combined accuracy and customer experience.
What should developers check before relying on it?
Start with the alias. Anything pointing at grok-voice-latest already runs 3.0, so test your agents now rather than waiting for a user complaint.
Then run a few real calls with your own accents, background noise and interruptions. Benchmarks use clean test sets, and your callers won’t.
Finally, watch your bills. The rate is flat, but better interruption handling can change how long calls last. A faster resolution means fewer billed minutes, and a looping bot means more. For a similar pricing story, see our look at Muse Wonder API pricing.
Frequently Asked Questions
How much does Grok Voice 3.0 cost?
It costs $0.08 per minute, which works out to $4.80 per hour. That’s the same price as Grok Voice 2.0.
Do I need to change my code to use Grok Voice 3.0?
Probably not. If your agent uses the grok-voice-latest alias, it was upgraded automatically. The endpoint, session parameters and events are unchanged.
Is Grok Voice 3.0 better than GPT-Live-1?
It depends on the test. xAI says it is cheaper and has 40% lower latency, but GPT-Live-1 scored higher on the τ-Voice support-call benchmark, 74.5% to 71.6%.
Is Grok Voice 2.0 still available?
It’s deprecated. xAI directs developers to migrate to Grok Voice 3.0.
Our take
Grok Voice 3.0 is an easy upgrade because it costs nothing extra and needs no code. The real test is your own call traffic, so run it side by side with whatever you use now before you commit. If you’re happy with your current voice model, there’s no rush, but if cost matters more than top support accuracy, this one deserves a trial.
Leave a Reply