Introduction

On October 14 (about 10:04 PT), SpaceXSI — the Grok team at x.ai — launched Grok Voice 3.0, a closed speech-to-speech model built for real-time voice agents. It is live today in the Voice Agent API and Voice Agent Builder at $0.08 per minute ($4.80 per hour), the same price as Grok Voice 2.0. There are no open weights.

The headline: Grok Voice 3.0 debuts at #1 on the Artificial Analysis Speech-to-Speech Index at 86.9, ahead of GPT-Live-1 (Astra, medium) at 83.2 and Gemini 3.8 Live Extended Thinking (High) at 82.6, while keeping time-to-first-audio under one second.

Primary source: the Introducing Grok Voice 3.0 announcement, plus the launch thread on X.

What Shipped

  • Model: Grok Voice 3.0 — successor to Grok Voice 2.0, a native speech-to-speech model (audio in, audio out) rather than a stitched STT → LLM → TTS chain.
  • Access: closed API over WebSockets at wss://api.x.ai/v1/realtime; Voice Agent Builder for no-code agents.
  • Migration: agents pointed at grok-voice-latest were switched to 3.0 automatically, with no code changes.
  • Price: $0.08/min ($4.80/hr of input audio), unchanged from 2.0.
  • Where it already runs: Tesla in-car voice, Grok Bot voice mode, and Starlink customer support, which SpaceXSI says handles about 20,000 calls a day with roughly 70% resolved with no human in the loop.

The Numbers

Artificial Analysis published its own evaluation (run with the Rex voice through the realtime API) alongside the launch:

  • Speech-to-Speech Index: 86.9, #1 — up 5.6 points from Grok Voice 2.0 (81.3) and 13.2 from Grok Voice 1.0 (73.7).
  • Big Bench Audio (speech reasoning): 98.5%, behind StepAudio 3 Realtime (99.7%), Qwen Audio 3.0 Realtime Plus (99.2%), and Qwen3.5 Omni Plus Realtime (98.7%), ahead of Gemini 3.8 Live (97.7%).
  • τ-Voice (agentic support calls with accents, noise, interruptions): 71.6%, #2 behind GPT-Live-1 (Astra, medium) at 74.5% — up from 56.5% on 2.0.
  • Speech Agent Arena: #3 on blind preference at Elo 1070, but #1 on task success at 94.8%.
  • Latency: 0.81 s average time to first audio on Big Bench Audio, about 40% faster than GPT-Live-1 (Astra) at 1.34 s. Grok Voice 2.0 was slightly quicker at 0.70 s.

SpaceXSI also cites 97% on Full Duplex Bench (pauses, turn-taking, barge-in) and the top accuracy score (74%) on ServiceNow's open EVA-Bench for enterprise voice agents.

One honest caveat: it is not a clean sweep. GPT-Live-1 (Astra) still edges it on τ-Voice, and two Gemini Live variants beat it on arena preference. The win is the combination of reasoning, task completion, and sub-second response at a mid-pack price.

Try It

The realtime API uses session events over WebSockets. A minimal Python setup from the docs:

import asyncio, json, os, websockets

async def main():
    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:
        await ws.send(json.dumps({
            "type": "session.update",
            "session": {
                "voice": "eve",
                "instructions": "You are a warm, concise support agent.",
                "turn_detection": {"type": "server_vad"},
            },
        }))
        async for msg in ws:
            event = json.loads(msg)
            if event["type"] == "response.output_audio.delta":
                pass  # play base64 audio chunk

asyncio.run(main())

Custom voices cloned from a reference clip (up to 120 seconds) through the Custom Voices API also work inside the realtime session, and the same voice roster covers every supported language.

Conclusion

If you build voice agents, Grok Voice 3.0 is a same-day closed upgrade with a real independent benchmark behind it and zero migration cost on grok-voice-latest. If you need open weights or local speech-to-speech, this is not that drop — keep watching the open omni stacks from StepFun and Qwen that still top raw speech reasoning. For everyone else, the fastest test is to point an existing agent at the realtime endpoint and run your own noisiest support call through it.

—Anabel ♡