What Is Grok Voice 3.0? 86.9 Index Score, $0.08/Min

Grok Voice 3.0 Is a Real-Time Speech-to-Speech Agent Model — 86.9 on Artificial Analysis at $0.08/Min

Grok Voice 3.0 is a real-time speech-to-speech model from SpaceXSI (formerly xAI). It listens to a caller, reasons, calls tools and answers in speech within one model, and it is built mainly for customer-support and phone agents. It was released on October 14, 2026, and it debuted at #1 on Artificial Analysis' Speech-to-Speech Index with a score of 86.9. It costs $0.08 per minute in SpaceXSI's Voice Agent API, which is the same price as Grok Voice 2.0. SpaceXSI says it already handles millions of conversations a day in Tesla vehicles and about 20,000 Starlink support calls a day.

Key Takeaways

• Status: Generally available since October 14, 2026, under the API model ID grok-voice-3.0. The grok-voice-latest alias now points to it.
• Rank: #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 at 82.6.
• Price: $0.08 per minute ($4.80 per hour) plus $0.004 per text input. This is unchanged from Grok Voice 2.0.
• Weak spot: On τ-Voice, the simulated support-call test, GPT-Live-1 still leads with 74.5% against Grok Voice 3.0's 71.6%.
• Latency: Under 1 second to first audio. SpaceXSI claims 40% lower latency than GPT-Live-1 on τ-Voice.
• Deployment: It runs Grok Bot's voice mode and Tesla in-car voice, and it powers Starlink call centers, where about 70% of calls are resolved with no human involved.

What Is Grok Voice 3.0?

Grok Voice 3.0 is the third generation of SpaceXSI's voice model and replaces Grok Voice 2.0, which is now deprecated. SpaceXSI's launch post calls it "the world's most intelligent speech-to-speech model." The post says it "sets a new record for resolving customer support calls."

Grok Voice 3.0 is a speech-to-speech model that hears, reasons, calls tools and talks back in one pass over a single WebSocket. Older voice bots chain speech-to-text, a text LLM and text-to-speech. Every handoff in that chain adds delay and loses tone. A native speech-to-speech model avoids those handoffs. That matters for barge-in, pauses and callers who change their mind halfway through a sentence.

The company behind it uses several names. SpaceXSI signs the docs and the launch post. Artificial Analysis lists the vendor as SpaceXAI, and many posts still say xAI. They are the same team, the same Grok family and the same api.x.ai endpoint. The official announcement on X went out on October 14, 2026.

Grok Voice 3.0 at a Glance

Attribute	Grok Voice 3.0
Developer	SpaceXSI (formerly xAI)
Type	Real-time speech-to-speech model for voice agents
Modality	Audio and text in, audio and text out (bidirectional streaming)
API model ID	grok-voice-3.0 (alias: grok-voice-latest)
Context window	Not yet confirmed
Pricing	$0.08/min ($4.80/hr) + $0.004 per text input
Time to first audio	Under 1 second (third-party reports: 0.81 s)
Availability	Voice Agent API, Voice Agent Builder, Grok Bot voice mode, Tesla vehicles
Release date	October 14, 2026
License	Proprietary hosted API; no open weights published
Parameter count	Not yet confirmed

How Grok Voice 3.0 Works / What Makes It Different

Grok Voice 3.0 runs on a WebSocket session. The client opens a connection to wss://api.x.ai/v1/realtime, chooses the model with a model query parameter, and configures the session with a session.update event. The Speech to Speech docs describe two authentication methods. Ephemeral Tokens are recommended for browsers and mobile apps. A raw API key is meant for server-side use only.

Five design choices stand out:

• Server VAD Turn Detection. With turn_detection set to server_vad, the model detects speech on its own and supports natural barge-in. On Full Duplex Bench, which measures pauses, turn-taking and interruptions, it scores 97.0%.
• Reasoning by Default. Developer coverage of the release notes reports that reasoning is on by default with effort set to "high". Developers can set it to "none" for speed.
• Alias Auto-Upgrade. Sessions pinned to grok-voice-latest moved to 3.0 automatically. Production teams that want a fixed model should pin grok-voice-3.0 and run their regression tests again.
• Agent Tooling In-Call. Agents can search the web and X, query uploaded collections, call functions and reach remote MCP servers while a call is in progress.
• SIP Telephony. The docs cover SIP Phone Calls, including transfers and keypad input. Reported audio formats are PCM, G.711 and Opus. Custom (cloned) voices are also supported.

The benchmark gain sits mostly in agentic support work. On τ-Voice, success rose from 56.5% on Grok Voice 2.0 to 71.6%, a 15.1-point jump. Speech reasoning was already near its ceiling, and Big Bench Audio rose only from 97.2% to 98.5%.

Benchmark (Artificial Analysis, Oct 13, 2026)	Grok Voice 3.0	Grok Voice 2.0	GPT-Live-1 (Astra, Medium)	Gemini 3.8 Live (Ext. Thinking, High)
AA Speech-to-Speech Index	86.9	81.3	83.2	82.6
Big Bench Audio	98.5%	97.2%	90.1%	97.7%
Speech Agent Arena Elo	1070	1011	1048	990
Task Success Rate	94.8%	94.6%	87.4%	89.1%
Full Duplex Bench	97.0%	95.1%	94.9%	91.9%
τ-Voice	71.6%	56.5%	74.5%	68.6%

Source: @cheatyyyy

Artificial Analysis' own launch thread confirms the #1 debut at 86.9. It notes that the model was tested through the Realtime API with the "Rex" voice. SpaceXSI also cites ServiceNow's EVA-Bench, which covers 200+ enterprise scenarios. Grok Voice 3.0 scores 74% accuracy there, the highest reported, but competitor figures were not published.

One number went the wrong way. Third-party writeups of the Artificial Analysis data put time to first audio at 0.81 seconds for Grok Voice 3.0, against 0.70 seconds for Grok Voice 2.0. A Chinese-language breakdown gives GPT-Live-1 at 1.34 seconds. It is still fast, but slightly slower than its predecessor.

What You Can Do With Grok Voice 3.0

SpaceXSI's own deployments show the intended workloads:

• Support call centers. According to the announcement, which The Verge highlighted, Starlink answers 20,000 calls a day with Grok Voice. A single agent verifies customers, troubleshoots hardware, issues credits and ships replacement equipment. About 70% of calls end with no human involved.
• Sales lines. One in four Starlink sales calls ends in a purchase before the caller hangs up. This figure is vendor-reported.
• In-car assistants. Tesla drivers use Connectors for calendar and email requests and Grok Bot for tasks like ordering coffee or booking dinner, all hands-free.
• Consumer voice chat. Grok Voice 3.0 powers Grok Bot's voice mode.
• Dubbing. In one community dubbing test, a lip-synced scene cost $0.012 with Grok Voice against $0.111 with GPT-audio, about 9x cheaper. That is a single unaudited test.

A short live demo video is also circulating. Teams that need only a narration voice and not a full agent may still prefer a cascaded setup, for example pairing a text model with Elevenlabs Text to Speech.

How Grok Voice 3.0 Compares

Grok Voice 3.0 leads Artificial Analysis' Speech-to-Speech Index at 86.9, which is 3.7 points ahead of GPT-Live-1 (Astra, Medium) at 83.2. The lead does not hold on every test, though.

Metric	Grok Voice 3.0	GPT-Live-1 (Astra, Medium)	Gemini 3.8 Live
AA index	86.9	83.2	82.6
τ-Voice	71.6%	74.5%	68.6%
Cost per task (Big Bench Audio)	$0.025	$0.030	Not yet confirmed
Time to first audio	0.81 s (reported)	1.34 s (reported)	Not yet confirmed
List price per minute	$0.08	Not yet confirmed	Not yet confirmed

GPT-Realtime-2.1 (High) costs $0.055 per task on the same Big Bench Audio run, more than double Grok Voice 3.0's figure. A third-party review notes that on Artificial Analysis' full leaderboard Grok Voice 3.0 ranks #3 on human preference, behind two Gemini entries. So "#1" means best average, not best on every axis. OpenAI responded within a day with GPT-Live-1.1, which adds new voices and is rolling out in ChatGPT, with API access "coming soon." For background, see the earlier GPT Live 1 full-duplex analysis.

Availability: How to Access Grok Voice 3.0

Grok Voice 3.0 is generally available through SpaceXSI's official channels:

1. Voice Agent API. Connect over WebSocket with model=grok-voice-3.0 or grok-voice-latest. The docs list tester apps for iOS, WebSocket, WebRTC and Twilio telephony.
2. Voice Agent Builder. SpaceXSI's no-code agent builder.
3. Grok Bot voice mode and Tesla vehicles for end users.

The endpoint, session parameters and events are unchanged from Grok Voice 2.0, so moving from 2.0 means changing one model string. SpaceXSI suggests rewriting old system prompts to be much shorter instead of copying them over verbatim. Grok Voice 2.0 is deprecated, and no retirement date has been published. A third-party review reports that the concurrency cap has moved from 10 calls to a spend-based ladder of up to 200, and that an EU region is now available.

What We Don't Know Yet

• Independent verification of deployment numbers. The 20,000 calls a day, the 70% resolution rate and the "millions" of Tesla conversations all come from SpaceXSI.
• Volume pricing. No tiers, discounts or per-voice price differences have been published.
• Voice coverage of the benchmarks. Artificial Analysis tested only the "Rex" voice.
• Multilingual quality. Community members have asked for side-by-side tests in languages such as Uzbek, Russian and German. None have been published.
• Context window, architecture and parameter count. None of these are disclosed.
• Leaderboard discrepancy. SpaceXSI's landing page claims "#1 Tau Voice Leaderboard," but Artificial Analysis' τ-Voice run places it #2.

Frequently Asked Questions

How much does Grok Voice 3.0 cost?

Grok Voice 3.0 costs $0.08 per minute ($4.80 per hour) plus $0.004 per text input in SpaceXSI's Voice Agent API. That is the same price as Grok Voice 2.0. Volume discounts and enterprise tiers have not been published.

Is Grok Voice 3.0 open source?

No. SpaceXSI offers Grok Voice 3.0 only as a hosted model through its API and products. It has published no model weights or open license.

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

Overall, yes: it scores 86.9 to 83.2 on Artificial Analysis' Speech-to-Speech Index. GPT-Live-1 (Astra, Medium) still leads the τ-Voice support-call benchmark, 74.5% to 71.6%. SpaceXSI says Grok Voice 3.0 reaches its τ-Voice result at 40% lower latency.

How do I use the Grok Voice 3.0 API?

Open a WebSocket to wss://api.x.ai/v1/realtime with model set to grok-voice-3.0 or grok-voice-latest. Use ephemeral tokens for client apps or an API key in the Authorization header on servers. It is also available in the Voice Agent Builder.

What is the difference between Grok Voice 3.0 and Grok Voice 2.0?

Grok Voice 3.0 scores 86.9 on the index versus 81.3, and it raises τ-Voice success from 56.5% to 71.6%. The API and the $0.08 per minute price stay the same. Third-party reports put its time to first audio at 0.81 seconds, slightly slower than 2.0's 0.70 seconds.

What languages does Grok Voice 3.0 support?

It supports more than 20 languages with automatic language detection, according to developer coverage of SpaceXSI's docs. SpaceXSI's voice landing page cites 25+ languages. Starlink uses it for support calls in multiple languages.

Who makes Grok Voice 3.0?

SpaceXSI, the company formerly known as xAI, makes Grok Voice 3.0. It also builds the Grok model family. Grok Voice 3.0 was released on October 14, 2026, and is served from the same api.x.ai endpoint as other Grok models.

What to Watch Next

For teams building phone agents, Grok Voice 3.0 is the highest-ranked voice model available at $0.08 per minute, though GPT-Live-1 still leads the τ-Voice support-call test. Three signals will shape updates to this page:

• Whether GPT-Live-1.1 reaches the API and changes the Artificial Analysis ranking.
• Independent production reports that test the 70% autonomous-resolution claim.
• A published retirement date for Grok Voice 2.0, which sets the migration deadline for teams that pinned the old model.

Try related models on Kie.ai: Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS.