xAI quietly turned Grok from a chatbot that could talk into a full voice platform. The roster went from five voices to 28, each one cast for a specific job: support, narration, commentary, advertising, education. All of them speak 25+ languages natively. Here’s the complete list, a sample of all 28, and how to pick one.

What actually changed

For most of Grok’s life, voice mode meant five options: Ara, Eve, Leo, Rex, and Sal. They were fine. They were also obviously the same voice engine wearing five different hats, and if you wanted an AI that could plausibly answer a support line or narrate an audiobook, you went to ElevenLabs.

The July 2026 update added 21 new flagship voices in one drop, and two more (Aurora and Liora) appeared on the roster shortly after without an announcement, which is very on-brand for xAI. That brings the documented total to 28 built-in voices. The original five were retrained at the same time for more natural pacing, phrasing, and emphasis, so even if you stick with Eve, she sounds different than she did in June.

Three things matter more than the headcount:

• They’re cast, not cloned. Each voice has a documented tone and a suggested use case. Carina is built for wellness and support. Zagan is built to play a villain. This is a departure from “here are twelve generically pleasant narrators.”
• Multilingual is native, not bolted on. Every voice handles 25+ languages without switching to a different model, so Luna sounds like Luna whether she’s explaining fractions in English or Mandarin.
• Sub-second response in realtime mode. The current model runs reasoning and still answers in under a second, which is the difference between a conversation and a walkie-talkie.

Listen first

Descriptions only get you so far. Here are eight of the voices, picked to cover the spread of the roster, so you can hear what “warm” and “cinematic” and “playful” actually mean before you read the table.

The full roster

These are the voice IDs exactly as you’d pass them to the API, with the tone and use-case labels xAI publishes for each one. Click any voice name to hear a sample of it.

The 23 flagship voices

Every voice name is a link. Click one to hear it.

The five originals (retrained)

Same here. Click a name to hear it.

How to pick one

The use-case tags are genuinely useful, but they overlap a lot. Nine voices claim “Assistant.” A few shortcuts that hold up in practice:

• Customer support: start with celeste or carina. Both are built to sound calm at the exact moment the caller isn’t. rigel is the pick if your support is more technical than emotional.
• Long-form narration or audiobooks: orion is the obvious cinematic choice; lux if you want something less theatrical that holds up over hours.
• Ads and brand work: altair for premium, perseus for trust, zenith for urgency.
• Education and explainers: cosmo for pace, luna for patience, lumen when you need both warmth and precision.
• Characters and commentary: zagan and sirius are the two with real personality. Everything else is a professional; these two are performers.

Where you can actually use these

In the Grok app

Open Grok on iOS, Android, or the web, go into voice settings, and pick from the list. If you don’t see the expanded options, update the app or clear the cache. The rollout is server-side and sometimes lags.

One caveat worth knowing: the consumer app exposes a much smaller set of voices with personality labels than the developer platform does. If you’re hunting for zagan or altair in the app and can’t find them, that’s why. The full 28 live on the API side.

Via the API

Three surfaces cover most needs:

• Text-to-Speech API: you send text, you get audio. Best for narration, ads, and anything pre-rendered.
• Realtime Voice Agent API: speech-to-speech, sub-second, handles interruptions and background noise. This is what you build a phone agent on. It supports direct SIP if you already have a phone system.
• Voice Agent Builder: a browser interface for building and deploying an agent with no code, in a couple of minutes.

A minimal TTS call looks like this:

curl https://api.x.ai/v1/audio/speech \
  -H "Authorization: Bearer $XAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "grok-voice-latest",
    "voice": "orion",
    "input": "Chapter one. The lights went out at 4:15."
  }' --output chapter-one.mp3

Swapping voices is a one-word change, which makes A/B testing tone almost free. Every sample above is the same pipeline with a different string. Check the xAI voice docs for the current endpoint shape before you ship, because the model names have already changed once.

Speech control tags

You can direct delivery inline rather than re-recording. Documented controls include pauses, whispers, and laughter, written directly into the text:

Thanks for calling. [pause] Let me pull that up for you.
<whisper>Between you and me, the sale ends tonight.</whisper>

This is the feature that closes most of the gap with dedicated TTS tools. A voice that can pause in the right place reads as human far more reliably than a voice that’s simply higher fidelity.

Voice cloning

You can clone a voice from roughly one to two minutes of audio. Availability is restricted: US-only, excluding Illinois (biometric privacy law), and limited to enterprise teams. If you’re a solo creator hoping to clone yourself this afternoon, that door isn’t open yet.

What it costs

• Text-to-Speech: $15.00 per 1M characters. Roughly 150,000 words, call it two novels, for fifteen dollars.
• Realtime speech-to-speech: starting at $0.05 per minute of audio, voice included.
• Telephony: about $0.01 per minute on top, on free-provisioned numbers.

For context, a support agent handling 1,000 five-minute calls a month runs around $250 in voice costs. That’s the number that makes this interesting to anyone who isn’t building a demo.

The honest caveats

Two things worth keeping in mind. First, xAI ships fast and documents late. Aurora and Liora showed up on the roster with no description and no announcement, and the older voice model was deprecated on a short runway. Pin your model version if you’re running anything in production.

Second, benchmark numbers on voice agents are published by the vendors running them, and the gap between a benchmark score and a caller who’s annoyed is wide. Run your own calls before you route real customers into one.

That said, the shape of the thing has changed. Twenty-eight cast voices, 25+ languages each, inline delivery control, and sub-second realtime for five cents a minute is a real platform, not a chatbot feature. If you tried Grok voice six months ago and shrugged, it’s worth another listen.

Sources

• xAI: 21 New Flagship Grok Voices
• xAI Docs: Voice guide and full voice table
• xAI: Grok Voice
• Basenor: Grok’s 21 New Voices Explained
• TestingCatalog: 21 new Grok Voices on the xAI Console