Grok Voice 3.0 Tops Artificial Analysis at 86.9 and Keeps $0.08 a Minute Pricing, but Still Trails GPT-Live-1 on Support Calls

SpaceXSI, the company most developers still call xAI, released Grok Voice 3.0 on 14 October 2026 and says it is the most intelligent speech to speech model available. Its launch post leans on one headline number: first place on the Artificial Analysis Speech to Speech Index with a score of 86.9.

The price did not move. Grok Voice 3.0 costs $0.08 per minute ($4.80 per hour) in the Voice Agent API, the same as Grok Voice 2.0, and any agent calling grok-voice-latest has already been switched over.

We checked the claims against the Artificial Analysis leaderboard and the developer docs. The top ranking holds up. A few of the supporting claims are narrower than the launch post makes them sound.

What SpaceXSI actually shipped

Grok Voice 3.0 is a native speech to speech model, so audio goes in and audio comes out without a separate transcription and text to speech chain. It runs on the same WebSocket endpoint as before, wss://api.x.ai/v1/realtime.

The Speech to Speech docs say the API, session parameters and events are unchanged from Grok Voice 2.0. The model list now reads grok-voice-latest as an alias for grok-voice-3.0, with grok-voice-2.0 marked deprecated. No shutdown date for 2.0 is listed yet.

One default changed in a way that matters for latency. The docs say reasoning is on by default with reasoning.effort set to “high”, and you have to set it to “none” to turn it off.

The branding is also in flux. The launch page and docs say SpaceXSI, while Artificial Analysis still lists the provider as SpaceXAI. We covered the SpaceXAI to SpaceXSI rename when Elon Musk first floated it. API hostnames and key names still say x.ai and XAI_API_KEY.

The benchmark numbers, checked

SpaceXSI published a comparison table sourced to Artificial Analysis and dated 13 October 2026. Here it is next to what the live leaderboard shows today.

Metric	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 (speech reasoning)	98.5%	97.2%	90.1%	97.7%
τ-Voice (support call tasks)	71.6%	56.5%	74.5%	68.6%
Arena preference Elo	1070	1011	1048 (1056 live)	990 (993 live)
Task success rate	94.8%	94.6%	87.4%	89.1%
Full Duplex Bench (conversation flow)	97.0%	95.1%	94.9%	91.9%
Time to first audio	0.81 s	0.70 s	1.34 s	1.35 s

Sources: SpaceXSI launch table and the Artificial Analysis summary table, read 15 Oct 2026. τ-Voice and latency figures for rivals come from Artificial Analysis.

The index win is real. At 86.9, Grok Voice 3.0 sits 3.7 points clear of GPT-Live-1 (Astra) and 4.3 points clear of Gemini 3.8 Live with extended thinking.

The index only ranks models that have all four components: speech reasoning, τ-Voice, arena preference and task success. Several models that beat Grok on individual tests are left out of the ranking because they are missing one piece.

Where the launch post stretches

On τ-Voice, the test closest to real customer support calls, SpaceXSI calls its 71.6% “comparable with GPT-Live-1 (Astra)”. Artificial Analysis lists GPT-Live-1 (Astra, medium) at 74.5%. That is close, but Grok is behind, and the leaderboard notes the GPT-Live-1 figure comes from a single trial.

The “40% lower latency” line does check out against that same rival. Grok’s 0.81 seconds to first audio is about 40% below GPT-Live-1’s 1.34 seconds. But Grok Voice 3.0 is slightly slower than Grok Voice 2.0, which measured 0.70 seconds.

Human preference is not a clean sweep either. The launch table compares against Gemini 3.8 Live with extended thinking at 990 Elo. On the same leaderboard, plain Gemini 3.8 Live scores 1088 and Gemini 3.1 Flash Live Minimal scores 1100, both above Grok’s 1070.

Single tests show the same pattern. StepAudio 3 Realtime and Qwen Audio 3.0 Realtime Plus both score higher than Grok on Big Bench Audio and Full Duplex Bench, and GPT-Live-1 (Sol, low) edges it on Full Duplex Bench at 97.3%. StepAudio lacks complete data, and the Qwen model falls to 69.5 on the index because it manages only 54.6% on τ-Voice.

Is it actually cheaper?

SpaceXSI says Grok Voice 3.0 costs less per task than GPT-Live-1 (Astra, Medium) and GPT-Realtime-2.1. Its chart puts Grok at $0.025 per task, against $0.03 and $0.055 for the two OpenAI models. The comparison leaves Google out.

Model	Artificial Analysis cost per hour of input audio (USD)
Gemini 3.8 Live	$0.84
Gemini 3.8 Live (Ext. Thinking, High)	$3.50
GPT-Live-1 (Sol, low)	$4.47
Grok Voice 3.0	$4.80
GPT-Live-1 (Astra, medium)	$5.83
GPT-Realtime-2.1 High	$10.75

Artificial Analysis measures this on a fixed 40-question Big Bench Audio subset, so treat it as a benchmark cost, not your bill. On that measure Grok is cheaper than most OpenAI options and pricier than Google’s Live models.

The SpaceXSI pricing table also lists a separate “$0.004 / text input” line next to the per-minute rate, without spelling out the unit. If your agent pushes a lot of text context into a session, check your first invoice before scaling up.

Tesla, Starlink and the ServiceNow quote

SpaceXSI says Grok Voice handles millions of conversations a day in Tesla vehicles and also powers Grok Bot’s voice mode. It says Starlink call centres take 20,000 customer calls a day on Grok Voice, with roughly 70% resolved without a human and one in four sales calls ending in a purchase.

Those are first-party numbers from deployments inside the same corporate family. No outside audit is cited.

The launch post also quotes Nitin Aggarwal, Senior Director of Product Management at ServiceNow, saying Grok Voice 3.0 “sets a new Pareto frontier on EVA-Bench”, ServiceNow’s open-source benchmark for enterprise voice agents. SpaceXSI puts its accuracy there at 74%, the highest score it reports.

Confirmed vs unconfirmed

Claim	Status
Available today in the Voice Agent API	Confirmed in launch post and docs
$0.08 per minute ($4.80 per hour), same as 2.0	Confirmed on pricing page
grok-voice-latest now points to 3.0	Confirmed in docs
#1 on AA Speech to Speech Index at 86.9	Confirmed on Artificial Analysis
97% Full Duplex Bench	Confirmed on Artificial Analysis
Comparable to GPT-Live-1 on τ-Voice	Partly: 71.6% vs 74.5%, Grok trails
Under one second latency	Confirmed: 0.81 s time to first audio, slower than 2.0
74% EVA-Bench accuracy	Vendor cited, with a ServiceNow quote
Starlink 70% resolution, Tesla volume	Unverified first-party claims
Grok Voice 2.0 shutdown date	Not announced

What it means for Indian developers

Hindi and Bengali are on the official supported language list, and the docs say the model detects the input language and can switch languages mid-conversation. That covers a lot of Indian support traffic, but not Tamil, Telugu, Marathi or Kannada, which are not listed.

At roughly ₹95 to the dollar, $0.08 a minute works out to about ₹7.60 a minute, or around ₹457 an hour. A support line doing 10,000 minutes a month would pay about $800 (roughly ₹76,000) before any text input charges.

The telephony support is practical for Indian call centres. The API takes G.711 μ-law and A-law audio natively and supports SIP, so it can plug into existing PBX setups without transcoding.

For Hinglish sales calls, compare it with local options. SquadStack recently opened Arth V2 with a benchmark built on Indian sales calls, which none of the Artificial Analysis tests are.

If you already run on OpenAI’s Realtime API, the docs say switching mostly means changing the base URL, the API key and the model name. OpenAI’s own voice lineup from DevDay 2026 is the obvious comparison.

Should you switch?

If you are on grok-voice-latest, you already did. Retest your prompts, because the docs tell 3.0 users to shorten system prompts and strip workarounds written for older models.

The same docs also advise pinning a versioned model name in production. That sits awkwardly with an alias that silently moved every agent to a new model on launch day. Pin grok-voice-3.0 once you have tested it.

If latency is your top priority, try reasoning.effort set to “none” and measure. The 0.81 second figure is Artificial Analysis’s measurement, and your network path to the API adds to it. Our earlier look at Gemma 4 on LiveKit at 354 ms shows how much a cascaded open setup can shave off.

FAQ

What does Grok Voice 3.0 cost?

$0.08 per minute, or $4.80 per hour, in the SpaceXSI Voice Agent API. That is the same as Grok Voice 2.0. The pricing page also lists a separate $0.004 text input line.

Is Grok Voice 3.0 really the top speech to speech model?

It ranks first on the Artificial Analysis Speech to Speech Index at 86.9. It does not lead every individual test: GPT-Live-1 (Astra) scores higher on τ-Voice, and some Gemini models score higher on human preference.

Do I need to change my code to use it?

No, if you use grok-voice-latest. If you pinned grok-voice-2.0, change the model to grok-voice-3.0. The endpoint and events are unchanged.

Does Grok Voice 3.0 support Hindi?

Yes. Hindi and Bengali are on the official supported language list in the Speech to Speech docs. Other Indian languages are not listed and may work with lower accuracy.

Is Grok Voice 2.0 being shut down?

It is marked deprecated in the docs, but no shutdown date has been published. Plan the move to grok-voice-3.0 now rather than waiting for a notice.