Grok Voice 3.0 tops speech benchmark with 0.81-second first audio

xAI has released Grok Voice 3.0, a native speech-to-speech model that ranks first on the Artificial Analysis Speech-to-Speech Index. It scored 86.9 on the composite index, 98.5% on audio reasoning, 94.8% task success, and 0.81 seconds to first audio. xAI kept the listed price at $0.08 per minute, giving voice-agent developers lower measured latency and stronger benchmark performance at the previous version’s rate.

Three tests behind the 86.9 score

Artificial Analysis calculates its index as an equal-weighted mean of three benchmarks, and models need valid results on all three to qualify:

• Speech Reasoning: Big Bench Audio measures whether a model can understand spoken information and reason about it.
• Conversational Dynamics: Full Duplex Bench evaluates pauses, turn-taking, and interruptions during live conversation.
• Agentic Performance: tau-Voice tests whether a voice agent can complete simulated customer-support tasks.

Grok Voice 3.0 improves on Grok Voice 2.0’s index score of 81.3 by 5.6 points. The original Grok Voice 1.0 scored 73.7, putting the total gain at 13.2 points in roughly six months.

Latency separates the leaders

Grok Voice 3.0 is the only configuration among the three index leaders to produce first audio in under one second. Its 0.81-second result is about 40% lower than the 1.34 to 1.35 seconds measured for the leading OpenAI and Google configurations. On tau-Voice, it improved from its predecessor’s 56.5% to 71.6%.

Configuration labels matter when comparing these results. Voice, reasoning level, and other runtime settings can affect quality, latency, and cost. A “Not reported” entry means the figure was absent from the comparison data shown here; index eligibility still requires a valid score for every component.

Where the benchmark lead narrows

Grok Voice 3.0 also scored 97.0% on Full Duplex Bench and 94.8% task success in Speech Agent Arena. The latter narrowly exceeds Grok Voice 2.0 at 94.6% and Gemini 3.8 Live at 93.2%.

Speech Agent Arena tracks task completion separately from blind-chat preference. Its Elo rating reflects head-to-head votes, with higher ratings indicating more wins. Grok Voice 3.0 placed third at 1070 Elo, behind Gemini 3.1 Flash Live Minimal at 1100 and Gemini 3.8 Live at 1088. The 94.8% Arena task-success result and the 71.6% tau-Voice score come from different evaluations and should be treated as separate metrics.

• Agentic tasks: GPT-Live-1 with Astra still leads Grok Voice 3.0 on tau-Voice, 74.5% to 71.6%.
• Raw latency: Grok Voice 2.0 remains faster to first audio at 0.70 seconds, compared with 0.81 seconds for 3.0.
• Input price: Grok costs $4.80 per input-audio hour, compared with $3.50 for the listed Gemini 3.8 configuration.
• Per-task cost: Artificial Analysis measured an average of $0.025 per Big Bench Audio task for Grok Voice 3.0, versus $0.030 for GPT-Live-1 with Astra at medium settings and $0.055 for GPT-Realtime-2.1 at high settings.

API access, aliases, and migration

Developers can access the model through the xAI Realtime API and Voice Agent API at $0.08 per minute, equivalent to $4.80 per hour. Deployments using the grok-voice-latest alias now resolve to version 3.0 automatically, so teams using that moving alias should include model upgrades in their regression-testing process.

Artificial Analysis evaluated the Rex voice, while the API exposes five personalities. Benchmark results for Rex should not be assumed to transfer exactly to every voice. Teams migrating from OpenAI’s Realtime API can retain the general WebSocket session pattern, then update the base URL, API key, model, and voice selection. Event schemas and interruption behavior should still be verified before production deployment.

A minimal Python connection requires the websockets package and an XAI_API_KEY environment variable:

import asyncio
import json
import os

import websockets


async def voice_agent():
    headers = {
        "Authorization": f"Bearer {os.environ['XAI_API_KEY']}"
    }

    async with websockets.connect(
        "wss://api.x.ai/v1/realtime?model=grok-voice-latest",
        additional_headers=headers,
    ) as ws:
        await ws.send(json.dumps({
            "type": "session.update",
            "session": {
                "voice": "rex",
                "instructions": "You are a helpful assistant.",
                "turn_detection": {
                    "type": "server_vad"
                }
            }
        }))

        async for raw_event in ws:
            event = json.loads(raw_event)
            print(event.get("type"), event)


asyncio.run(voice_agent())

This connection configures Rex, enables server-side voice activity detection, and logs incoming events. A production client must also stream microphone audio, play returned audio, handle interruptions and cancellation, reconnect after failures, and restrict any tools the agent can call.

Best fit: tool-using voice agents

The combination of 0.81-second first audio, 97.0% full-duplex performance, and strong task completion makes Grok Voice 3.0 a candidate for phone support and other voice workflows that call tools during a conversation. xAI also advertises support for dozens of languages and mid-conversation language switching, which broadens the range of call-center and multilingual applications worth testing.

Production evaluations should cover more than the published averages. Teams should measure end-to-end and tail latency, telephony codec quality, accents, background noise, interruption recovery, tool-call accuracy, escalation behavior, and total cost across both successful and failed tasks.

Native speech-to-speech systems compete with pipelines that connect automatic speech recognition, a language model, and text-to-speech synthesis. Big Bench Audio scores have moved from the low 90s a year ago to above 97% for three vendors, making latency, conversational control, task completion, and human preference more useful differentiators. Grok Voice 3.0 leads the composite benchmark through strong results across those dimensions, while its third-place preference rating and higher price than Gemini leave clear criteria for production testing.