xAI Grok Voice 3.0 API Teardown: $0.08/Min Pricing, 120s Voice Cloning & Fixing the -latest Alias Bug

Quick answer (checked 18 October 2026): Grok Voice 3.0 launched on 14 October 2026 in xAI's Voice Agent API at $0.08 per minute ($4.80 per hour), the same price as Grok Voice 2.0. Any app using the grok-voice-latest alias was moved to 3.0 automatically. That's why some voice agents feel slower: on Artificial Analysis, average time to first audio went from 0.70 s (2.0) to 0.81 s (3.0), and 3.0 runs with reasoning effort high by default. To control it, pin grok-voice-3.0 instead of the alias, and set reasoning.effort to "none" for simple, latency-sensitive flows. Custom voices can be cloned from a reference clip of up to 120 seconds.

Here's a situation many voice-agent builders ran into this week. Nothing in your code changed. No deploy went out. But callers say your agent pauses a little longer before it speaks, and your latency dashboard agrees.

If your WebSocket URL ends in ?model=grok-voice-latest, that's the explanation. On 14 October, xAI (now branded SpaceXAI) released Grok Voice 3.0 and pointed the alias at it. Its launch post says agents on the alias "have already been switched over automatically, with no changes needed".

For most apps that's an upgrade: 3.0 scores higher on almost every benchmark xAI published. But it's a slightly slower and smarter model by default, and an alias that changes under you is a production risk. This guide covers what changed, what it costs, and how to take back control.

A note on sources: everything here comes from xAI's launch post and developer docs and from Artificial Analysis's speech-to-speech leaderboard, all linked at the end. I haven't run my own latency tests on 3.0, so the figures are theirs.

What Grok Voice 3.0 is

Grok Voice 3.0 is a speech-to-speech model. Audio goes in and audio comes out over one real-time connection, with no separate speech-to-text, LLM and text-to-speech steps chained together. It's built for voice agents: support lines, phone bots, in-app assistants.

According to xAI's launch post, it also powers Grok's voice mode, calls in Tesla vehicles and Starlink call centers.

Key facts from the docs:

xAI's docs say 3.0 uses the same endpoint, session parameters and events as 2.0. So migration is a one-word change, which is exactly why the automatic alias switch was possible.

The benchmarks, read carefully

xAI's launch post compares Grok Voice 3.0 with its predecessor, OpenAI's GPT-Live-1 and Google's Gemini 3.8 Live. Most figures come from Artificial Analysis:

Two things stand out:

1. The biggest jump is on agent tasks. τ-Voice rose from 56.5% to 71.6%. That benchmark measures whether a voice agent completes real multi-step tasks, which is the thing most builders care about.
2. It doesn't win everything. GPT-Live-1 (Astra) still scores higher on τ-Voice, 74.5% vs 71.6%. xAI's argument is that it gets close with much lower latency.

xAI also lists cost per task on Big Bench Audio: $0.025 for Grok Voice 3.0, $0.030 for GPT-Live-1 (Astra, Medium) and $0.055 for GPT-Realtime-2.1 (High).

The latency change: 0.70 s → 0.81 s

Artificial Analysis measures time to first audio (TTFA), the average wait before the model starts speaking, on its Big Bench Audio set:

So both of these are true at once:

• Grok Voice 3.0 is still faster than every OpenAI and Google model on the list, which is the basis of xAI's "under one second" claim. On τ-Voice, xAI says it has 40% lower latency than GPT-Live-1 (Astra) at comparable accuracy.
• It's about 0.11 s slower than Grok Voice 2.0 on average. If your agent was tuned around 2.0's timing, for example with tight barge-in or silence thresholds, you might notice.

The likely reason is the default reasoning setting. In the docs, reasoning.effort defaults to "high" and accepts only "high" or "none". A model that thinks before it speaks takes a bit longer to start. xAI hasn't published a TTFA figure for 3.0 with reasoning set to "none", so measure it yourself (step 4 below).

Fix 1: pin the model version

Aliases are handy in development and risky in production. For anything customer-facing, use the versioned name:

# Before: follows whatever xAI points the alias at
MODEL = "grok-voice-latest"

# After: stays on 3.0 until you decide to change it
MODEL = "grok-voice-3.0"

url = f"wss://api.x.ai/v1/realtime?model={MODEL}"

Should you pin back to grok-voice-2.0? Only as a short emergency measure. The docs mark 2.0 as deprecated. xAI hasn't published a shutdown date on the pages I checked, but deprecated models do get removed. Pinning 3.0 and tuning it is the long-term fix.

Fix 2: turn reasoning off for simple flows

Many voice flows don't need the model to reason: confirming an appointment, reading an order status, routing a call. For those, send a session.update with reasoning off:

{
  "type": "session.update",
  "session": {
    "instructions": "You are a short, friendly order-status assistant.",
    "voice": "eve",
    "reasoning": { "effort": "none" },
    "turn_detection": { "type": "server_vad" },
    "audio": {
      "input":  { "format": { "type": "audio/pcm", "rate": 24000 } },
      "output": { "format": { "type": "audio/pcm", "rate": 24000 } }
    }
  }
}

Keep "high" for flows where the agent has to work something out, such as troubleshooting, multi-step booking or anything with several tool calls. That's where 3.0's τ-Voice gain lives.

A practical pattern is to decide per agent, not per app. Your "reception" agent that routes calls can run with reasoning off, while your "support" agent keeps it on.

Fix 3: retune your turn-taking

If you tuned silence and interruption timing for 2.0, revisit it:

• turn_detection.type: "server_vad" lets the server detect when the caller stops talking. Set it to null only if you manage turns yourself.
• turn_detection.idle_timeout_ms controls how long a session waits on silence (default null).
• audio.output.speed ranges from 0.7 to 1.5 (default 1.0). Slightly faster speech can make the whole exchange feel snappier, even when TTFA is unchanged.

Step 4: measure your own latency

Don't rely on leaderboard averages. Your prompts, tools and network decide your real numbers. A minimal approach:

1. Record the time when the caller's turn ends (when VAD reports speech stopped, or when you commit the audio buffer).
2. Record the time when the first audio chunk of the response arrives.
3. The difference is your time to first audio. Log it per session with the model name and reasoning setting.
4. Run 50 calls with reasoning.effort: "high" and 50 with "none" on the same script, and compare the medians and the slowest 10%.
5. Check task success too. A faster agent that gets more answers wrong isn't an improvement.

The pricing, worked out

At $0.08 per minute, the maths is simple:

Because the price didn't change from 2.0, the upgrade doesn't change your bill directly. What can change it is call length: if 3.0 resolves tasks more often on the first try, calls get shorter and you pay less. If it talks longer, you pay more. Track average minutes per resolved call before and after.

Custom voices from a 120-second clip

xAI's Custom Voices API lets you clone a voice from a reference recording and use it in the Voice Agent API by passing the returned voice_id as the voice value. The rules, from the docs:

• Clip length: up to 120 seconds. Longer clips are rejected with a 400. xAI recommends 90–120 seconds.
• Best format: WAV (uncompressed PCM), 24 kHz, 16-bit, mono. MP3, FLAC, OGG, Opus, M4A, AAC, MKV and MP4 are accepted, but lossy formats can add artefacts.
• Endpoint: POST /v1/custom-voices as multipart/form-data, with the audio as file. There are matching list, get, update, download and delete endpoints.
• Who can use the API: creating voices through the API is limited to Enterprise plans. Other teams get a 403 and can create voices in the console instead, up to 30 free custom voices per team.
• Region: Custom Voices is currently only available in the United States, except Illinois.

Only clone voices you have the right to use. That means your own, a voice actor's under contract, or a brand voice your company owns. Cloning someone's voice without their consent can break the law in many places, and it's an easy way to lose a customer's trust.

Tips for a good reference clip:

1. Record in a quiet room with a decent USB microphone, not a laptop mic.
2. Speak naturally, in the tone the agent will use: calm for support, upbeat for sales.
3. Avoid background music and long pauses.
4. Export as WAV, mono, 24 kHz.

Keep your API key off the client

Browsers can't set WebSocket headers, and you should never ship your API key in front-end code anyway. xAI's answer is ephemeral tokens:

1. Your server calls POST https://api.x.ai/v1/realtime/client_secrets with your API key and a short expiry, for example {"expires_after": {"seconds": 300}}.
2. It passes the short-lived token to the browser or app.
3. The client connects with the token in the WebSocket protocol as xai-client-secret.<token>.

If the token leaks, it expires on its own in minutes.

Migrating from OpenAI's Realtime API

xAI says the Voice Agent API is compatible with the OpenAI Realtime API specification, and its docs include migration examples for OpenAI SDK users. In practice, the main changes are the endpoint URL, the API key and the model name. Test tool calls and voice names carefully, though, since each provider has its own voices and tool options (x_search is xAI-specific, for example).

Common mistakes to avoid

• Running production on -latest. Use grok-voice-3.0 and upgrade on purpose.
• Pinning grok-voice-2.0 and forgetting. It's deprecated. Treat it as a stopgap.
• Leaving reasoning on everywhere. Simple flows rarely need it, and it costs latency.
• Judging the model on TTFA alone. Check task success and call length too.
• Uploading a noisy clip for a custom voice. Garbage in, garbage voice out.
• Putting the API key in the browser. Use ephemeral tokens.
• Cloning a voice without consent. Get written permission first.

Final thoughts

Grok Voice 3.0 is a real upgrade, especially for agents that have to finish tasks, and it costs the same as 2.0. The 0.11-second slowdown is small, and still well ahead of OpenAI's and Google's voice models on Artificial Analysis's numbers. What actually hurt some teams this week wasn't the model. It was finding out from their users that the alias had moved.

So pin your model, decide reasoning per agent, and measure your own latency. Then you choose when to upgrade, not the alias.

Frequently asked questions

How much does Grok Voice 3.0 cost?

$0.08 per minute ($4.80 per hour) in xAI's Voice Agent API, the same as Grok Voice 2.0.

Why did my Grok voice agent get slower?

If you use grok-voice-latest, it moved to Grok Voice 3.0 on 14 October 2026. Artificial Analysis measures average time to first audio at 0.81 s for 3.0 vs 0.70 s for 2.0, and 3.0 defaults to high reasoning effort.

How do I reduce Grok Voice 3.0 latency?

Set reasoning.effort to "none" in session.update for simple flows, pin grok-voice-3.0, and retune your turn detection. Measure time to first audio before and after.

Is Grok Voice 3.0 better than GPT-Live-1?

On xAI's published numbers it scores higher on the AA Speech-to-Speech Index (86.9 vs 83.2), Big Bench Audio and task success, and starts speaking faster. GPT-Live-1 (Astra) scores higher on τ-Voice (74.5% vs 71.6%).

How long can a Grok custom voice reference clip be?

Up to 120 seconds; xAI recommends 90–120 seconds of clean WAV audio at 24 kHz mono. Longer clips are rejected.

Can I still use Grok Voice 2.0?

It's still listed but marked deprecated. Pin it only briefly while you test 3.0.

Is the Grok Voice Agent API compatible with OpenAI's Realtime API?

xAI says it's compatible with the OpenAI Realtime API specification, and its docs include migration examples for OpenAI SDK users.

Sources

• SpaceXAI (xAI): Introducing Grok Voice 3.0 (14 October 2026)
• xAI Docs: Speech to Speech / Voice Agent API
• xAI Docs: Custom Voices
• xAI Docs: Ephemeral tokens
• Artificial Analysis: Speech-to-Speech leaderboard (time to first audio)

Related on VNHAX: What is SpaceXAI? · xAI roadmap: Grok 4.8 and 4.9 · Grok 4.7 vs Grok 4.6 · ElevenLabs v4 emotive voice guide