Grok Voice 3.0 Takes the Top Spot on Speech, at Eight Cents a Minute
Quick verdict
SpaceXAI released Grok Voice 3.0 and called it its most intelligent speech-to-speech model. For once the independent numbers back the marketing. Artificial Analysis put it first on its Speech-to-Speech Index at 86.9, ahead of OpenAI's GPT-Live-1 Astra at 83.2. It answers in about eight tenths of a second and costs $0.08 per minute of input. That combination, top of an outside leaderboard plus a clear per-minute price, is what separates this from the usual voice-model demo reel.
What actually shipped
Grok Voice 3.0 is a speech-to-speech model, meaning it takes audio in and speaks audio back rather than routing through a text step in the middle. That design is what keeps the response latency low. The headline specs from Artificial Analysis and SpaceXAI:
- First on the Artificial Analysis Speech-to-Speech Index at 86.9, ahead of GPT-Live-1 Astra at 83.2
- 71.6% on Tau Voice and 98.5% on Big Bench Audio
- About 0.81 seconds to first audio
- $0.08 per minute, which works out to $4.80 per hour of input audio
The sub-second response time is the number that matters most for anything conversational. Voice assistants live or die on whether the reply feels like a conversation or a help-desk queue, and 0.81 seconds is on the right side of that line.
How it compares
The interesting part is that an independent evaluator, not the vendor, put Grok Voice 3.0 on top. Vendor benchmarks are marketing until someone else runs the test, and here Artificial Analysis ran it.
| Model | Speech-to-Speech Index |
|---|---|
| Grok Voice 3.0 | 86.9 |
| GPT-Live-1 Astra | 83.2 |
OpenAI has been the name to beat in realtime voice since it shipped its full-duplex API, so an outside leaderboard with xAI's model in front is a real shift. Our writeup on GPT-Live-1 and full-duplex voice covers what OpenAI built and where it was ahead, which is the thing Grok Voice 3.0 now claims to have passed.
Why it matters
Voice is turning into a real product surface rather than a party trick, and the price is why. At $0.08 per minute, a voice agent that handles a five-minute call costs about forty cents to run, which is cheap enough to put in front of customers instead of keeping it in a demo. For anyone building with voice, the choice is no longer "OpenAI or wait," and competition on a leaderboard usually drags prices down and quality up for everyone. If you want to keep several of these models on hand instead of betting on one, our guide to running multiple AI models in one place walks through why that beats a single subscription.
Video: Grok Voice 3.0 explained
A walkthrough of what xAI shipped, how the speech-to-speech setup works, and where it lands against the other voice models.
FAQ
Is Grok Voice 3.0 really better than OpenAI's voice model?
On the Artificial Analysis Speech-to-Speech Index, yes: it scores 86.9 against 83.2 for GPT-Live-1 Astra. That is one benchmark from one evaluator, so it is a lead rather than a rout, but it is an independent one rather than xAI's own numbers.
What does Grok Voice 3.0 cost?
$0.08 per minute of input audio, or about $4.80 per hour. For a voice agent, that puts a typical short call in the range of a few cents to half a dollar depending on length.
What is a speech-to-speech model?
It takes audio in and returns audio out without converting to text in the middle, which is what keeps the response time low. Grok Voice 3.0 answers in about 0.81 seconds, close enough to feel like a live conversation.
Sources
- @SpaceXSI - Grok Voice 3.0 launch
- Artificial Analysis - #1 on the Speech-to-Speech Index, scores, latency and pricing
Further reading
Try all the models mentioned in this article
Admix gives you GPT-5, Claude, Gemini, and 350+ AI models in one app. Compare responses side by side. Try it for 7 days for $1.99.
Try Admix for $1.99