Grok Voice 3.0: What It Is, Benchmarks, Pricing and How to Use It
xAI released Grok Voice 3.0 on 14 Oct 2026: a speech-to-speech model ranked #1 on the Artificial Analysis index at the same $0.08 per minute as 2.0. Benchmarks, what changed, and how to try it.
Last updated:
xAI shipped Grok Voice 3.0 on 14 October 2026. It is the speech-to-speech model behind the Grok app's voice mode and the Grok Voice API, and xAI's launch post puts it at the top of the main third-party speech-to-speech leaderboard. The price did not move: $0.08 per minute, the same as Grok Voice 2.0, which is now marked deprecated.
This page collects what is actually known about the model, what changed from 2.0, how it is priced, and how to get it into your own app. If you only want to hear the voices, the text to speech tool on the homepage uses the same 28 voices.
What Grok Voice 3.0 is
Grok Voice 3.0 is a speech-to-speech model: it takes audio in and produces audio out in one pass, instead of chaining speech recognition, a text model and a text to speech voice. That is what lets it do full-duplex conversation (talking and listening at once, handling interruptions), keep sub-second latency, and carry tone from the user's voice into its answer.
xAI runs it as the engine of:
- the Grok app's voice mode on iOS, Android and the web;
- the Grok Voice API for developers building voice agents (support lines, sales calls, in-app assistants), including phone calls over SIP;
- its own operations: xAI says Grok Voice handles customer-support calls for Tesla and for SpaceX's Starlink, at about 20,000 calls a day, with roughly 70% resolved without a human.
The model id in the API is grok-voice-3.0; the alias grok-voice-latest always points at the newest version.
Benchmarks xAI reports
The numbers below come from xAI's launch post (x.ai/news/grok-voice-3). They are the vendor's own figures, with the comparisons the vendor chose, so treat them as a starting point and listen for yourself.
| Benchmark | Grok Voice 3.0 | Grok Voice 2.0 | Others cited by xAI |
|---|---|---|---|
| Artificial Analysis Speech-to-Speech Index | 86.9 (#1) | 81.3 | GPT-Live-1 83.2 · Gemini 3.8 Live 82.6 |
| Big Bench Audio (reasoning over audio) | 98.5% | — | — |
| Speech Agent Arena (Elo) | 1070 | — | — |
| Task success rate | 94.8% | — | — |
| Full Duplex Bench | 97.0% | — | — |
| τ-Voice (tool-using voice agents) | 71.6% | — | — |
| EVA-Bench (expressiveness) | 74% | — | — |
Two of these matter most if you are choosing a model for an agent: τ-Voice, which scores whether a voice agent completes multi-step tasks with tools, and Full Duplex Bench, which scores turn-taking and barge-in. Expressiveness (EVA-Bench) is what you hear in the voices themselves.
What changed from Grok Voice 2.0
- Higher scores across the board: 86.9 vs 81.3 on the Artificial Analysis index is the headline, and xAI frames 3.0 as the first speech-to-speech model to pass the previous leaders from OpenAI and Google.
- Reasoning you can switch: sessions accept a
reasoning.effortofhigh(default) ornone. Turn it off for simple, latency-sensitive flows; leave it on for multi-step requests. - Same 28 voices: 3.0 uses the lineup xAI expanded in July 2026 (21 new flagship voices joining the original five), so the API lists 28 built-in voices, from the original Ara, Eve, Leo, Rex and Sal to Altair, Atlas, Aurora, Carina, Castor, Celeste, Cosmo, Helios, Helix, Iris, Kepler, Liora, Lumen, Luna, Lux, Naksh, Orion, Perseus, Rigel, Sirius, Ursa, Zagan and Zenith. All speak all supported languages. See the full voice list.
- Same price: $0.08 per minute, $0.004 per text input. Grok Voice 2.0 stays available at the same rate but is deprecated.
Grok Voice 3.0 pricing
| Item | Price |
|---|---|
| Speech-to-speech session (grok-voice-3.0) | $0.08 per minute ($4.80 per hour) |
| Text injected into a session (instructions, tool results) | $0.004 per text input |
| Text to speech (same voices, no conversation) | $15.00 per 1M characters |
| Speech to text | $0.10 per hour (REST), $0.20 per hour (streaming) |
Our pricing calculator turns minutes, characters and hours into a monthly figure.
How to try Grok Voice 3.0
Without code. Open the Grok app and start voice mode; pick a voice in the app's settings. For the voices alone, use the free Grok text to speech tool here: type anything, choose a voice, download the MP3.
With code. Get an API key at console.x.ai, then:
- For a server-side agent, open a Realtime WebSocket session with
"model": "grok-voice-latest"and the voice, instructions and tools you want. xAI's console also has a no-code agent builder. - For a browser or mobile client, first call
POST /v1/realtime/client_secretsfrom your server to mint an ephemeral token (default 10 minutes, maximum 1 hour), and connect the client with that token instead of your API key. - For phone calls, provision a number with
POST /v2/phone-numbers(xAI-provisioned or your own SIP trunk) and route it to an agent or a webhook.
The Grok voice agent guide goes through the session setup, and the Grok Voice API guide covers the text to speech and speech to text endpoints that sit next to it.
Grok Voice 3.0 vs. Grok text to speech
People search for both and mix them up, so to be precise:
- Grok Voice 3.0 = the conversational model. Audio in, audio out, reasoning and tools in between. Billed per minute.
- Grok text to speech =
POST /v1/tts. Text in, audio out, no conversation. Billed per character. Same 28 voices, plus custom cloned voices.
If your product reads things aloud (narration, notifications, accessibility), you want text to speech. If it talks with people, you want Grok Voice 3.0.