Home › Guides › ElevenLabs vs Cartesia
HONEST REVIEW · CURRENT AS OF SEPTEMBER 2026
ElevenLabs vs Cartesia (2026): Which AI Voice Wins — Tested and Scored Language by Language
TL;DR verdict
English narration: ElevenLabs Eleven v4. It leads the independent Artificial Analysis Provider Voice Arena at Elo 1,319 (read September 28, 2026), and our own hands-on v4 samples below show why — expressive, direction-following narration.
Non-English and multilingual: Cartesia Sonic 3.6. The Sonic family leads eight of the nine non-English leaderboards tracked by Artificial Analysis (reported September 22, 2026) — that’s independent data, not our ears.
Real-time agents: no crowned winner. ElevenLabs claims ~100ms median latency for v4 Turbo; Cartesia claims sub-90ms time-to-first-audio for Sonic 3.6. Both are vendor claims, not our measurements.
The honesty line, up front: we have ElevenLabs API access but no Cartesia access. Our hands-on tests on this page cover Eleven v4 only. Everything about Sonic 3.6 comes from independent third-party leaderboards and is labeled as such. We will not claim to have heard what we have not heard.
Last checked: September 30, 2026.
What changed in September 2026
This comparison flipped this week, so it’s worth saying plainly what’s new. On September 28, 2026, ElevenLabs launched Eleven v4 and Eleven v4 Turbo: a rebuilt text-to-speech architecture with inline direction tags for emotion, pacing, reactions, and sound effects, 90+ language coverage (up from roughly 70), and a Turbo variant aimed at real-time conversation. On launch day, Artificial Analysis’ Provider Voice Arena ranked Eleven v4 #1 in English at Elo 1,319 — ahead of Cartesia’s Sonic 3.6 at 1,276.
That displaced a genuine incumbent. Cartesia’s Sonic 3.6 had topped both Artificial Analysis speech leaderboards in mid-August (1,283 provider-voice Elo), and the Sonic family still leads eight of the nine non-English boards. So this isn’t a story about one winner. It’s a story about a map that got redrawn along language lines — and that’s exactly how we’ll judge it.
How we tested (and what we didn’t)
Our testing policy is simple: publish what we ran, label what we didn’t. Here’s the split for this page:
- Hands-on, our own ears: ElevenLabs
eleven_v4— two stock demo voices on the same script, plus a direction-tag demo of v4’s headline feature. Generated September 30, 2026, with default settings. Playable below. - Independent third-party data: everything about Cartesia Sonic 3.6 — Artificial Analysis Provider Voice Arena (English board, read September 28, 2026) and the nine non-English leaderboards (reported September 22, 2026). Cited with dates, never presented as our listening.
- Vendor claims, labeled: v4 Turbo’s ~100ms latency, Sonic 3.6’s sub-90ms time-to-first-audio, 10-second instant cloning. Real numbers from the companies; not our measurements. We say so every time.
- Not tested: Sonic 3.6 audio (no Cartesia API access), real-time agent round-trips for either vendor, and voice cloning fidelity. When we run those, they’ll be published with audio, settings, and dates.
Why this matters: most “X vs Y” posts mash vendor marketing and third-party scores into one blended verdict. We keep the three piles separate — our ears, independent votes, vendor claims — so you know exactly what’s behind every sentence below.
Our hands-on v4 listening test
Voice A
Same script, default settings. Guess which stock demo voice this is before you peek.
Voice B
Identical script and settings, different stock demo voice. Listen first, reveal second.
Click to reveal: which voice is which?
- Voice A = Sarah (voice ID
EXAVITQu4vr4xnSDxMaL) — filev4-voicea-sarah.mp3 - Voice B = Roger (voice ID
CwhRBWXzGAHq8TQ4Fs17) — filev4-voiceb-roger.mp3
Both were generated 2026-09-30 on eleven_v4 with identical default settings. The point of the test: v4 holds delivery quality across voices on the same script — acronyms (Sonic, Elo), proper nouns, and pacing all land cleanly. Which one you prefer for your channel is the verdict that matters; neither is the locked Ear Tested narrator voice.
v4’s headline feature: inline direction
The v4 feature we actually wanted to poke at is inline direction — plain-text tags like [excited] and [whisper] that steer delivery mid-line. So we generated one take to check the vendor claim with our own ears:
Direction demo
One take, three directions, no regeneration tricks. Listen for whether the delivery actually shifts.
Our finding: the direction tags work, first attempt, no cherry-picking — the read genuinely shifts energy between the three segments. For YouTube narration, this is the practical difference between v4 and the older Multilingual v2: you can punch up a line without re-recording it or stacking a second generation. Cartesia’s Sonic has its own inline expression tags (e.g. [laughter]), which we have not tested hands-on — that comparison stays untested until we can run both.
The independent data: language by language
For the Cartesia side, we go to the independent record: Artificial Analysis’ arenas, where blind listeners vote between two clips. Here’s the board as read September 28–30, 2026:
| Rank | Model | Elo | Listed price / 1M chars |
|---|---|---|---|
| 1 | Eleven v4 (ElevenLabs) | 1,319 | $80 (intro $22 until Oct 12) |
| 2 | Sonic 3.6 (Cartesia) | 1,276 | $49 |
| 3 | Gemini 3.8 Flash TTS (Google) | 1,267 | $16.50 |
Elo from 1,674 blind samples on the English provider board. The arena uses each provider’s native voices and default configuration — v4’s 43-point lead is a clear first place on that board, and ElevenLabs also reports ~75% preference in its own head-to-head tests (vendor measurement, not the arena’s).
Two honest caveats from the same source, because the fine print is the whole story:
- The controlled-voice board disagrees. When every model is forced onto the same cloned voices (stripping out the voice-catalog advantage), Eleven v4 ranks second at 1,157 Elo behind Alibaba’s Qwen-Audio-3.1-TTS-Plus at 1,178. That suggests part of v4’s provider-board lead comes from ElevenLabs’ voices and defaults, not the synthesis engine alone. Worth knowing if you’ll clone a fixed voice.
- English is not the world. On the nine non-English boards Artificial Analysis added (Spanish, French, German, Portuguese, Japanese, Hindi, Mandarin Chinese, Arabic, Vietnamese — 100,000+ native-speaker votes), the Sonic family leads eight of nine as reported September 22, 2026. The exact winning Sonic version varies in Portuguese; Inworld takes Mandarin. If your audience isn’t English-first, the English board is the wrong map.
So the language-by-language picture is genuinely split: v4 wins English preference votes; Sonic 3.6 wins the non-English map. That’s not hedging — it’s what the independent data says, and it drives the verdict below.
Cost math for YouTube creators
Our rule of thumb stands: about 1,000 characters of script ≈ one minute of finished narration. Here’s how the two vendors price a creator’s actual workload:
- Eleven v4 (API): list price $80 per million characters, with an introductory rate of $22 per million running through October 12, 2026 — per reporting on ElevenLabs’ pricing page, read September 28, 2026. v4 Turbo lists at $40 per million, intro $11. Note: ElevenLabs’ own pricing page could not be loaded directly during this run, so these are dated third-party reads of it, not a live check.
- Sonic 3.6 (API): Artificial Analysis lists $49 per million characters. Cartesia’s own pricing page (read September 30, 2026) prices in credits — 1 credit = 1 character for Sonic TTS — with overages at $65/1M (Pro), $45/1M (Startup), $38/1M (Scale). The $49 figure sits squarely in that range.
- Creator math on subscriptions: ElevenLabs’ $6/month Starter tier covers ~30,000 characters — about three 10-minute videos a month (~$2.00 per video). Cartesia is credit-plan based ($5 for 100K credits at the low end), built for API/agent workloads rather than a creator subscription ladder.
What this means: for a YouTuber narrating a few videos a week, ElevenLabs’ subscription ladder is the simpler buy — and the free tier exists for non-commercial testing. For API-volume or agent workloads, Sonic 3.6’s ~$49/1M undercuts v4’s $80/1M list by nearly 40% (less during the intro window). Price follows the use case, same as everything else on this page. And the licensing wall is identical: free tiers are non-commercial — monetized videos need a paid plan from day one.
The verdict, by use case
No single winner — here’s the split we’d actually act on:
- English YouTube narration & character work → ElevenLabs v4. #1 on the independent English board, and our own hands-on samples show expressive, direction-following delivery. The inline direction tags are a real workflow upgrade for narration. Our tests + independent data.
- Non-English or multilingual channels → Cartesia Sonic 3.6. Leads eight of nine non-English leaderboards on 100,000+ native-speaker votes. If your audience is Spanish-, French-, Hindi-, or Arabic-first, the English ranking is irrelevant to you. Independent data.
- Real-time voice agents → untested, test both. v4 Turbo’s ~100ms and Sonic’s sub-90ms are both vendor claims; end-to-end latency depends on your network, model, and playback. Sonic’s credit pricing is friendlier at agent volume. Vendor claims + pricing data.
- Fixed cloned voice → compare directly. The controlled-voice board puts Alibaba’s Qwen model ahead of v4 — and we haven’t run our own cloning comparison, so we won’t name a winner. Clone the same voice on both and A/B it blind. Independent data; cloning untested by us.
- Bulk, low-scrutiny generation → price it out. Sonic 3.6 at ~$49/1M beats v4’s $80/1M list for volume work; both beat subscription pricing at scale. Pricing data.
Our recommendation for Ear Tested readers: if you narrate English YouTube videos, v4 is the default we’d start from — the free tier costs nothing to test with your own scripts, and Starter at $6/month covers about three 10-minute videos. If you publish in other languages, start with Sonic 3.6 instead. Either way: generate your actual scripts, listen blind, and let your ears break the tie — that’s what this site is for.
Free guide: Use AI Like the Top 1%
Voiceover is one chapter of a bigger system. Our free 21-page guide covers the AI workflows top creators actually use — scripting, titles, editing, repurposing, and batching. Get the free guide.
FAQ
Is ElevenLabs or Cartesia better for YouTube narration in 2026?
For English YouTube narration, the evidence favors ElevenLabs Eleven v4: it leads the Artificial Analysis Provider Voice Arena at Elo 1,319 (read September 28, 2026), and our own hands-on v4 samples on this page back up the narration quality. For non-English narration, Cartesia’s Sonic 3.6 is the stronger pick — it leads eight of nine non-English leaderboards.
Did Ear Tested run its own Cartesia Sonic 3.6 listening tests?
No. We have ElevenLabs API access but no Cartesia access, so our hands-on tests cover Eleven v4 only, with real generated samples. Everything about Sonic 3.6 here comes from independent third-party data — Artificial Analysis leaderboards read September 22–28, 2026 — and is labeled as such. We will not claim to have heard what we have not heard.
How much does ElevenLabs v4 cost compared to Cartesia Sonic 3.6?
Eleven v4’s API list price is $80 per million characters, with an introductory rate of $22 per million through October 12, 2026 (per reporting on ElevenLabs’ pricing page, read September 28, 2026). Artificial Analysis lists Sonic 3.6 at $49 per million, and Cartesia’s own pricing page quotes overages of $45–$65 per million characters by tier (read September 30, 2026).
Which is better for real-time voice agents: ElevenLabs v4 Turbo or Cartesia Sonic 3.6?
Both are built for speed and we have not tested either live, so we name no winner. ElevenLabs claims ~100ms median inference latency for v4 Turbo and Cartesia claims sub-90ms time-to-first-audio for Sonic 3.6 — both are vendor claims, not our measurements. For conversational agents, test both under your real network conditions.
Is Cartesia better than ElevenLabs for non-English videos?
According to independent data, yes: Cartesia’s Sonic family leads eight of the nine non-English language boards tracked by Artificial Analysis (Spanish, French, German, Portuguese, Japanese, Hindi, Arabic, Vietnamese), as reported September 22, 2026. Eleven v4 supports 90+ languages, but independent non-English preference votes still favor Sonic 3.6.
Can I use ElevenLabs’ or Cartesia’s free tier for monetized YouTube videos?
No — ElevenLabs’ free tier is licensed for non-commercial use only; monetized videos need a paid plan starting at the $6/month Starter tier. Cartesia’s plans are credit-based and its free tier is meant for evaluation, not commercial production. Budget for a paid plan from day one if you monetize.