Grok Voice 3.0: What's New, API Pricing and How to Use It

Grok Voice 3.0 is xAI's speech-to-speech model for real-time voice agents, announced on October 14, 2026. You talk, it listens and answers in its own voice over one WebSocket connection, with no separate speech-to-text or text-to-speech step.

If you build voice agents, three things matter: how well it listens and reasons, what a minute costs, and how much code changes. For Grok Voice 3.0 the answers are first place on Artificial Analysis's speech-to-speech index (86.9, as xAI reports it), $0.08 a minute, and no change at all if you already call grok-voice-latest.

Short answer

• Same price, better model. Grok Voice 3.0 replaces Grok Voice 2.0 at the same $0.08 a minute, on the same endpoint, with the same session events.
• It beats Grok Voice 2.0 on every score xAI compared: 86.9 vs 81.3 on the speech-to-speech index, and 97.0% vs 95.1% on handling pauses and interruptions.
• Connect over WebSocket at wss://api.x.ai/v1/realtime. The API follows OpenAI's Realtime format, so OpenAI SDKs work with a new base URL.
• It speaks 20+ languages with 28 built-in voices, and it can call tools such as web search and your own functions mid-conversation.

What is Grok Voice 3.0?

It is a speech-to-speech model: audio goes in, audio comes out, and the model reasons in between. xAI describes it as a "speech-to-speech model for real-time voice agents". It handles millions of conversations a day in Tesla cars and powers the voice mode of Grok Bot, and Starlink's call centers answer 20,000 calls a day with Grok Voice.

The model ID is grok-voice-3.0. The alias grok-voice-latest has pointed to it since launch day, so any session that uses the alias was upgraded automatically. Grok Voice 2.0 is now marked deprecated.

What changed from Grok Voice 2.0?

Every score xAI compared went up, while the price and the API stayed the same. These numbers come from xAI's announcement, which credits Artificial Analysis's tests from October 13, 2026.

The same table puts GPT-Live-1 (Astra) at 83.2 on the index and Gemini 3.8 Live at 82.6, and Grok Voice 3.0 leads both on all five rows.

Full Duplex Bench measures how naturally a model handles pauses, turn-taking and interruptions, which is where callers notice a voice agent most. xAI also says the model answers in under one second.

The model reasons before it speaks. Reasoning is on by default, and you can turn it off per session when speed matters more than depth. A session can also call tools while it talks: file search, web search, X search, remote MCP servers and your own functions.

How much does the Grok Voice API cost?

Grok Voice 3.0 costs $0.08 per minute of audio, or $4.80 per hour, according to xAI's pricing page. That page also lists "$0.004 / text input" for the same model without saying what the unit is.

If you only need one direction, xAI prices those separately: text to speech is $15 per million characters, and speech to text is $0.10 an hour (or $0.20 an hour streaming).

How do you connect to it?

You open a WebSocket to wss://api.x.ai/v1/realtime, pass the model as a query parameter, configure the session, then stream audio both ways. This is xAI's own Python quick start, copied from its docs:

import asyncio
import json
import os
import websockets

async def voice_agent():
    async with websockets.connect(
        "wss://api.x.ai/v1/realtime?model=grok-voice-latest",
        additional_headers={"Authorization": f"Bearer {os.environ['XAI_API_KEY']}"}
    ) as ws:
        # Configure session
        await ws.send(json.dumps({
            "type": "session.update",
            "session": {
                "voice": "eve",
                "instructions": "You are a helpful assistant.",
                "turn_detection": {"type": "server_vad"}
            }
        }))

        # Send a text message
        await ws.send(json.dumps({
            "type": "conversation.item.create",
            "item": {"type": "message", "role": "user",
                     "content": [{"type": "input_text", "text": "Hello!"}]}
        }))
        await ws.send(json.dumps({"type": "response.create"}))

        # Receive audio/text responses
        async for msg in ws:
            event = json.loads(msg)
            print(f"Event: {event['type']}")

asyncio.run(voice_agent())

To pin the new model instead of the alias, use model=grok-voice-3.0. A few details decide whether your first call works:

Which voices and languages does it have?

The docs list 28 built-in voices, with eve as the default, and every voice can speak every supported language. The speech-to-speech API supports more than 20 languages. It detects the caller's language, answers in it, and can switch languages mid-conversation.

You can also clone a voice from a reference clip of up to two minutes with xAI's Custom Voices API. Each team can create up to 30 custom voices for free, but the feature is only available in the United States, except Illinois.

xAI says voice audio is processed in real time and never stored or used for training.

Can you use Grok Voice 3.0 through OmniaKey?

Not today. OmniaKey does not offer Grok Voice, so you call it with an xAI API key. If you also use Grok's text and image models, Grok 4.7 and Grok Imagine 2.0 run on OmniaKey with one key, next to Claude, GPT and Gemini. Our Grok Imagine API guide shows the image side.

Common questions

Is Grok Voice 3.0 free?

The API is paid: xAI's pricing page lists $0.08 per minute of audio. If you only want to talk to Grok, the Grok apps on the web, iOS and Android have a hands-free voice mode.

Do I need to change my code from Grok Voice 2.0?

No. The endpoint, session settings and events are the same, and grok-voice-latest already points to 3.0. Only code pinned to grok-voice-2.0 needs the new model name.

Does Grok Voice 3.0 work with the OpenAI SDK?

Yes. xAI's migration guide says an app built on OpenAI's Realtime API needs three changes: the base URL (https://api.x.ai/v1), the API key and the model name.

How fast does Grok Voice 3.0 respond?

xAI says it answers in under one second. Its announcement also says its τ-Voice score of 71.6% is comparable to GPT-Live-1 (Astra) at 40% lower latency.

Sources

• xAI announcement: Grok Voice 3.0 (October 14, 2026)
• xAI docs: speech to speech and voice overview
• xAI pricing, rate limits and release notes
• xAI Custom Voices

All checked on October 19, 2026.