Grok Voice 3.0 vs elevenlabs: $0.08/min Speech-to-Speech

Grok Voice 3.0 vs ElevenLabs: $0.08 a Minute and an 86.9 Speech-to-Speech Index Debut

Grok Voice 3.0 is a fully benchmarked speech-to-speech model. It costs $0.08 per minute and debuted #1 on the Artificial Analysis Speech-to-Speech Index with 86.9. ElevenLabs has no published head-to-head numbers against it, so the choice comes down to architecture. Pick Grok Voice 3.0 if you want one real-time model that listens, reasons and speaks. Look at ElevenLabs if you need a separate text-to-speech layer that you control yourself. SpaceXSI (formerly xAI) released Grok Voice 3.0 on October 14, 2026. Every Grok number below is vendor- or benchmark-reported. Every ElevenLabs cell is marked unverified until someone publishes comparable data.

Key Takeaways

• Grok Voice 3.0 has the data on its side. It scores 86.9 on the Artificial Analysis Speech-to-Speech Index, 98.5% on Big Bench Audio, 97.0% on Full Duplex Bench and 71.6% on τ-Voice. None of the sources reviewed include ElevenLabs in these tests.
• Grok's price is public and unchanged: $0.08 per minute ($4.80 per hour) plus $0.004 per text input, the same as Grok Voice 2.0. No ElevenLabs price appears in the sources reviewed.
• The two products are different shapes. Grok Voice 3.0 is a Single-Pass Speech-to-Speech model. The ElevenLabs models in this site's catalog are text-to-speech and text-to-dialogue products.
• Grok trails on one test. On τ-Voice, the support-call benchmark, Grok Voice 3.0 (71.6%) trails GPT-Live-1 (74.5%). It is also 0.11 seconds slower to first audio than Grok Voice 2.0.
• Head-to-head data is still thin. Treat any "Grok beats ElevenLabs" or "ElevenLabs beats Grok" claim as opinion until a shared test exists.

Grok Voice 3.0 vs ElevenLabs at a Glance

Dimension	Grok Voice 3.0	ElevenLabs
Product type	Speech-to-speech model (grok-voice-3.0), streamed over WebSocket	Catalogued as text-to-speech / text-to-dialogue (ElevenLabs V3, ElevenLabs Text to Speech)
Price	$0.08/min ($4.80/hr) + $0.004 per text input (official)	Unverified (no public number in the sources reviewed)
Artificial Analysis S2S Index	86.9, #1 at debut (Artificial Analysis, Oct 13, 2026)	Unverified (no public number yet)
Time to first audio	0.81 s average (reported)	Unverified (no public number yet)
Support-call benchmark (τ-Voice)	71.6%	Unverified (no public number yet)
Telephony / tools	SIP calls, transfers, keypad input, web and X search, remote MCP	Not yet confirmed

Data asymmetry: every Grok Voice 3.0 figure comes from the official launch post or Artificial Analysis's same-day evaluation, and no independent hands-on test has been published. The ElevenLabs column is empty because no source reviewed here published ElevenLabs data on these dimensions.

Architecture: One Model vs a Voice Layer

This is the dimension that decides the comparison.

Grok Voice 3.0 is a Single-Pass Speech-to-Speech model. Older voice bots chain three systems together: speech-to-text, then a text model, then text-to-speech. One builder's explainer puts it this way: every handoff between them "adds delay and also loses the tone." A speech-to-speech model does all three jobs in one pass. That lets it pick up that a caller is annoyed and answer before they repeat themselves.

The ElevenLabs models listed in this site's catalog are classed as text-to-speech and text-to-dialogue. In a real-time agent, that kind of model is the final voice stage of a pipeline. Something else handles listening and reasoning. Neither design is better in the abstract. A cascaded pipeline lets you swap each part. A single-pass model removes the handoffs.

According to the official speech-to-speech docs, Grok Voice 3.0 streams audio and text in both directions over wss://api.x.ai/v1/realtime. You authenticate with ephemeral tokens on the client side or an API key on the server side. Developer coverage of the release documents more:

• Audio and voices: PCM, G.711 and Opus audio, cloned voices, and automatic language detection.
• Server-Side Barge-In: turn detection set to server_vad lets callers interrupt naturally.
• Reasoning-On-by-Default: reasoning effort defaults to "high", and "none" is the only other value.
• Phone calls: calls over SIP, with transfers and keypad input.

xAI's own API page also lists separate text-to-speech pricing at $15.00 per 1M characters. So SpaceXSI sells both shapes. The comparison with ElevenLabs is clearest on the speech-to-speech side, which is where Grok Voice 3.0 lives.

If you want to hear the text-to-speech shape side by side, ElevenLabs V3 is in this site's catalog as a text-to-dialogue model. Pair it with a separate listener and reasoner to see how the cascaded approach compares in your own stack.

Benchmarks: What Is Measured, and Against Whom

Grok Voice 3.0's benchmark record is detailed, but ElevenLabs is absent from it. The official launch post publishes this table, credited to Artificial Analysis and dated October 13, 2026:

Benchmark	Grok Voice 3.0	Grok Voice 2.0	GPT-Live-1 (Astra, Medium)	Gemini 3.8 Live (Ext. Thinking, High)	ElevenLabs
AA Speech-to-Speech Index	86.9	81.3	83.2	82.6	Not yet confirmed
Big Bench Audio	98.5%	97.2%	90.1%	97.7%	Not yet confirmed
Speech Agent Arena Elo	1070	1011	1048	990	Not yet confirmed
Task Success Rate	94.8%	94.6%	87.4%	89.1%	Not yet confirmed
Full Duplex Bench	97.0%	95.1%	94.9%	91.9%	Not yet confirmed
τ-Voice	71.6%	56.5%	74.5%	68.6%	Not yet confirmed

Source: @testingcatalog

Artificial Analysis's launch evaluation confirms the #1 debut at 86.9. It also notes the test ran through the Realtime API using the "Rex" voice. On ServiceNow's EVA-Bench, SpaceXSI reports the highest accuracy, 74%, but publishes no competitor figures.

Two caveats matter. First, "#1" means the best average, not the best at everything: GPT-Live-1 still leads τ-Voice by 2.9 points. Second, one builder review points out that Grok ranks #3 on human preference in the full Artificial Analysis leaderboard, behind two Gemini entries.

The fair summary: Grok Voice 3.0 is the best-documented speech-to-speech model available, and ElevenLabs simply has no entry in these comparisons yet.

For context on the model Grok edges past on the index, see our earlier analysis of GPT Live 1's full-duplex design.

Pricing and Latency

Grok Voice 3.0 costs $0.08 per minute, plus $0.004 per text input. That works out to $4.80 per hour and matches Grok Voice 2.0. The grok-voice-latest alias moved to 3.0 automatically, and Grok Voice 2.0 is now deprecated. An older version of xAI's API pricing page, captured in August 2026, lists speech-to-speech "starting at $0.05 / min." The launch materials quote $0.08 per minute for Grok Voice 3.0 specifically. Treat $0.08 as the current figure.

On cost per task, Artificial Analysis's Big Bench Audio run puts Grok Voice 3.0 at $0.025. That compares with $0.030 for GPT-Live-1 (Astra, Medium) and $0.055 for GPT-Realtime-2.1 (High). One community dubbing test put Grok Voice at $0.012 against $0.111 for GPT audio on the same clip. That is a single, informal comparison, not a measurement.

On latency, Grok Voice 3.0 reportedly starts speaking after 0.81 seconds on average, against 1.34 seconds for GPT-Live-1. That is about 40% faster. It is slower than Grok Voice 2.0's 0.70 seconds, though.

ElevenLabs pricing and latency are unverified, with no public number in the sources reviewed. Any "X times cheaper than ElevenLabs" claim you see online lacks a public, like-for-like basis right now.

Deployment Evidence

Grok Voice 3.0 already runs at scale, according to SpaceXSI. It handles millions of conversations a day in Tesla vehicles and powers Grok Bot's voice mode. The Verge highlighted one Starlink detail from the announcement: Starlink call centers answer 20,000 calls a day with Grok Voice, about 70% resolved with no human, and one in four sales calls ends in a purchase. These are vendor claims with no independent confirmation.

Third-party reviews also report operational changes since 2.0: concurrency now scales up to 200 calls on a spend ladder, and there is a new EU region. The language count varies by source, with "20+" in the docs coverage and "25+" on the landing page. No comparable deployment figures for ElevenLabs appear in the sources reviewed.

Which One Should You Use?

Choose Grok Voice 3.0 if:

• You are building a real-time phone or in-app support agent that has to take actions. It has documented SIP, tool calling, remote MCP and barge-in.
• You want published numbers to justify the choice: 86.9 on the index, 94.8% task success, $0.08 per minute.
• You want one model to replace a three-stage listen-think-speak pipeline.

Choose ElevenLabs if:

• You need a separate voice-generation layer, such as narration or dialogue from text, rather than a conversational agent.
• You already run your own speech recognition and language model and only want to swap the voice stage.
• You have tested its voices on your own content. No public benchmark against Grok Voice 3.0 exists to decide this for you.

Test both if: voice quality on your scripts is the deciding factor. Human-preference data comparing the two hasn't been published.

Frequently Asked Questions

Is Grok Voice 3.0 better than ElevenLabs?

Nobody has published a head-to-head test. Grok Voice 3.0 debuted #1 on the Artificial Analysis Speech-to-Speech Index with 86.9, ahead of GPT-Live-1 (83.2) and Gemini 3.8 Live (82.6). ElevenLabs doesn't appear in that comparison.

Is Grok Voice 3.0 cheaper than ElevenLabs?

Public sources can't answer that yet. Grok Voice 3.0 costs $0.08 per minute ($4.80 per hour) plus $0.004 per text input. No ElevenLabs price appears in the sources reviewed for this page.

What is the difference between Grok Voice 3.0 and ElevenLabs?

Grok Voice 3.0 is a speech-to-speech model: one model hears, reasons, calls tools and replies in voice over a WebSocket. ElevenLabs models are catalogued as text-to-speech and text-to-dialogue products. In a voice agent, they would sit in a pipeline alongside separate speech-recognition and language models.

Can Grok Voice 3.0 replace ElevenLabs for voice agents?

For real-time agents, Grok Voice 3.0 is built to handle the whole conversation in one model, with SIP calls, barge-in and tool calling documented. Nobody has tested in public whether it matches ElevenLabs on voice quality for narration or dubbing.

How fast is Grok Voice 3.0 compared to ElevenLabs?

Grok Voice 3.0 reportedly reaches first audio in 0.81 seconds on average, against 1.34 seconds for GPT-Live-1. No comparable ElevenLabs latency figure has been published in the sources reviewed.

Does Grok Voice 3.0 support custom or cloned voices like ElevenLabs?

Yes. Developer coverage of the docs lists cloned voices, PCM, G.711 and Opus audio, and automatic language detection. Nobody has measured how its cloned voices compare in quality with ElevenLabs voices.

What to Watch Next

Three things will change this page. The first is whether Artificial Analysis or another evaluator adds ElevenLabs to the same speech-to-speech or human-preference leaderboard. The second is independent latency and cost-per-resolved-call tests from builders putting Grok Voice 3.0 on live phone lines. The third is whether SpaceXSI sets a retirement date for Grok Voice 2.0, or OpenAI's newly announced GPT-Live-1.1 shifts the τ-Voice lead. For a sense of where open text-to-speech is heading, see our Fish Audio Drama 3 deep dive.

Try related models on Kie.ai: Gemini 3.8 Flash TTS, Gemini 3.8 Flash-Lite TTS and Gemini 3.1 Flash TTS.