SpaceXAI just released Grok Voice 3.0. Grok Voice 3.0 is a SpaceXAI speech-to-speech model that listens, reasons, calls tools and speaks. It now ranks first on the Artificial Analysis Speech-to-Speech Index with a score of 86.9. For developers, the endpoint, events and $0.08 per minute price are unchanged from Grok Voice 2.0.
TL;DR
- Runs on: SpaceXAI’s hosted Voice Agent API only. No open weights, no local or quantized variants.
- Performance: Highest Index score of any model Artificial Analysis currently ranks.
- Best: 86.9 on the AA Speech-to-Speech Index, ranked #1.
- Worst: 71.6% on τ-Voice, behind GPT-Live-1 (Astra, medium) at 74.5%.
- Bottom line:
- Best: top composite score at 0.81 s time to first audio.
- Worst: not the leader on agentic task resolution or raw speech reasoning.
What is Grok Voice 3.0?
Grok Voice 3.0 is SpaceXAI’s flagship realtime voice model, exposed as grok-voice-3.0 on the /v1/realtime WebSocket endpoint. The grok-voice-latest alias already points to it, so existing agents switched automatically.
SpaceXAI has not published parameter counts or architecture details. Reasoning is on by default (reasoning.effort: "high") and can be set to "none" for faster turns. Server-side voice activity detection handles barge-in. Tools run inside the session: collections search, web search, X search, remote MCP servers and custom functions.
How does Grok Voice 3.0 perform on benchmarks?
The AA Speech-to-Speech Index is an equal-weighted average of 4 components. Those are Big Bench Audio reasoning, τ-Voice agentic performance, Speech Agent Arena preference and task success rate.
Grok Voice 3.0 scores 86.9. GPT-Live-1 (Astra, medium) scores 83.2, Gemini 3.8 Live Extended Thinking 82.6. Grok Voice 2.0 scored 81.3.
Component results are strong, not uniformly first:
- Big Bench Audio: 98.5%, ranked 4th. StepAudio 3 Realtime leads at 99.7%.
- Full Duplex Bench: 97.0%, ranked 4th. StepAudio 3 Realtime leads at 98.9%.
- Task success rate: 94.8%, up from 94.6% for Grok Voice 2.0.
- Speech Agent Arena Elo: 1070, the highest among the 4 models compared here.
- τ-Voice: 71.6%, up from 56.5%, but below GPT-Live-1 (Astra, medium) at 74.5%.
ServiceNow’s EVA-Bench scores enterprise voice agents on accuracy and experience. SpaceXAI reports a top accuracy score of 74%.
Is Grok Voice 3.0 fast and cheap?
Artificial Analysis measures 0.81 s time to first audio. GPT-Live-1 (Astra, medium) measures 1.34 s, so SpaceXAI’s “40% lower latency” claim checks out. Grok Voice 2.0 was faster, though, at 0.70 s.
On cost, SpaceXAI cites Artificial Analysis cost per Big Bench Audio task. Grok Voice 3.0 costs $0.025 per task, versus $0.030 for GPT-Live-1 (Astra, medium) and $0.055 for GPT-Realtime-2.1 (High). On AA’s cost per hour of input audio, Gemini 3.8 Live Extended Thinking is cheaper: $3.50 versus $4.80.
List prices are not directly comparable. GPT-Live-1 costs $0.05 per minute, but backend model and tool usage is billed separately. Gemini 3.8 Live bills audio per minute plus text tokens.
</button><button class="btn" id="gvPlay">Auto-play</button></div>
</div>
<div class="pnl" id="p2">
<div class="chips" id="gvChips"></div>
<div id="gvBars"></div>
<p class="note" id="gvNote"></p>
</div>
<div class="pnl" id="p3">
<svg viewBox="0 0 640 330" width="100%" role="img" aria-label="τ-Voice success rate versus time to first audio">
<rect x="0" y="0" width="640" height="330" fill="transparent"/>
<g stroke="#2a2a2a">
<line x1="70" y1="30" x2="70" y2="280"/><line x1="70" y1="280" x2="610" y2="280"/>
<line x1="70" y1="196.7" x2="610" y2="196.7" stroke-dasharray="3 4"/><line x1="70" y1="113.3" x2="610" y2="113.3" stroke-dasharray="3 4"/>
<line x1="70" y1="30" x2="610" y2="30" stroke-dasharray="3 4"/>
</g>
<g fill="#9a9a9a" font-size="11">
<text x="62" y="284" text-anchor="end">50%</text><text x="62" y="201" text-anchor="end">60%</text><text x="62" y="117" text-anchor="end">70%</text><text x="62" y="34" text-anchor="end">80%</text>
<text x="70" y="298" text-anchor="middle">0.5 s</text><text x="205" y="298" text-anchor="middle">0.75 s</text><text x="340" y="298" text-anchor="middle">1.0 s</text><text x="475" y="298" text-anchor="middle">1.25 s</text><text x="610" y="298" text-anchor="middle">1.5 s</text>
<text x="340" y="320" text-anchor="middle">Time to first audio (lower is faster)</text>
<text x="16" y="155" text-anchor="middle" transform="rotate(-90 16 155)">τ-Voice success</text>
</g>
<text x="84" y="50" fill="#ffffff" font-size="12" font-weight="700">
better</text>
<g id="gvDots"></g>
</svg>
<p class="note">Source: Artificial Analysis, accessed October 14, 2026. Grok Voice 3.0 trails GPT-Live-1 (Astra, medium) by 2.9 points on τ-Voice but reaches first audio 0.53 s sooner.</p>
</div>
<div class="pnl" id="p4">
<div class="kv">
<div class="k"><b>$0.08/min</b><span>$4.80/hr, same as Grok Voice 2.0</span></div>
<div class="k"><b>0.81 s</b><span>time to first audio (AA)</span></div>
<div class="k"><b>28</b><span>built-in voices, plus cloning</span></div>
<div class="k"><b>20</b><span>listed locales, auto-detected</span></div>
<div class="k"><b>5</b><span>tool types: files, web, X, MCP, functions</span></div>
<div class="k"><b>SIP</b><span>phone and contact-center routing</span></div>
</div>
<pre><code>wss://api.x.ai/v1/realtime?model=grok-voice-3.0
{"type": "session.update",
"session": {
"voice": "eve",
"reasoning": {"effort": "high"},
"turn_detection": {"type": "server_vad"},
"tools": [{"type": "web_search"}]
}}</code></pre>
<p class="note">The <code>grok-voice-latest</code> alias already points to <code>grok-voice-3.0</code>. Pin the versioned name in production. Grok Voice 2.0 is deprecated.</p>
</div>
<div class="ft"><span>Data: <a href="https://x.ai/news/grok-voice-3" target="_blank" rel="noopener">SpaceXAI</a>, <a href="https://artificialanalysis.ai/speech-to-speech" target="_blank" rel="noopener">Artificial Analysis</a>, <a href="https://docs.x.ai/developers/model-capabilities/audio/speech-to-speech" target="_blank" rel="noopener">SpaceXAI docs</a></span><span class="brand">© Marktechpost</span></div>
</div>
<script>
(function(){
var R=document.getElementById('mtp-gv3');
function post(){try{parent.postMessage({mtpEmbed:'gv3',height:R.offsetHeight+40},'*')}catch(e){}}
R.querySelectorAll('.tab').forEach(function(t){t.addEventListener('click',function(){
R.querySelectorAll('.tab').forEach(function(x){x.classList.remove('on')});
R.querySelectorAll('.pnl').forEach(function(x){x.classList.remove('on')});
t.classList.add('on');document.getElementById(t.dataset.p).classList.add('on');
if(t.dataset.p==='p2')drawBars();if(t.dataset.p==='p3')drawDots();setTimeout(post,60);});});
/* wave */
var w=document.getElementById('gvWave');for(var i=0;i<48;i++){var b=document.createElement('i');b.style.animationDelay=(i*0.045)+'s';w.appendChild(b);}
/* pipeline */
var S=[
['Caller speaks','Audio streams over a WebSocket (or SIP for phone calls) as PCM at 8 to 48 kHz, G.711 or Opus.'],
['Turn detection','Server-side voice activity detection decides when the caller has finished, and handles barge-in when they interrupt.'],
['Reasoning','Reasoning is on by default with reasoning.effort set to "high". Set it to "none" for faster, simpler turns.'],
['Tools','The agent can call collections search, web search, X search, remote MCP servers and your own functions mid-call.'],
['Reply streams','Audio deltas stream back as soon as they are ready. Artificial Analysis measures 0.81 s to first audio.']];
var F=document.getElementById('gvFlow'),cur=0,timer=null;
S.forEach(function(s,i){var b=document.createElement('button');b.className='st';b.innerHTML='<div class="n">STEP '+(i+1)+'</div><div c