ElevenLabs v4 vs v4 Turbo: Is the #1 TTS Model Actually #1?

ElevenLabs says Eleven v4 is ranked #1 by Artificial Analysis, and the claim holds on one board. On the Provider Voice Arena (captured 16 October 2026), Eleven v4 Turbo scores 1330 Elo and Eleven v4 scores 1329, so the "#1" is a two-way tie inside the error bars, not a flagship beating its rivals alone. On two other independent measures, Controlled Voice and Artificial Analysis's pronunciation benchmark, neither model leads.

This teardown covers what the arena measures, how it is scored, where ElevenLabs' own numbers diverge from independent ones, and which variant to pick.

Which is ranked higher, Eleven v4 or Eleven v4 Turbo?

They are statistically tied. As of 16 October 2026, Artificial Analysis lists Eleven v4 Turbo at 1330±18 (95% interval 1312 to 1348, 1,821 samples) and Eleven v4 at 1329±17 (95% interval 1312 to 1346, 2,121 samples). Both carry the rank range 1-2.

Snapshots of this board move. In its launch post, Artificial Analysis gave Eleven v4 an Elo of 1,319 (+19/-19) across 1,674 appearances. According to eesel, an early-October explainer showed Turbo at 1334 and v4 at 1321, a 13-point gap that has since shrunk to one point. Pick any single snapshot and you can tell either story.

What does the Provider Voice Arena measure?

It measures blind listener preference for each model's own stock voices, on English prompts. Per the methodology page, each comparison plays two clips generated from the same text, with voices matched on gender and accent. Prompts span Customer Service, Entertainment, Knowledge Sharing and Assistants, and include challenge material such as alphanumeric strings.

The voice set is small. Artificial Analysis selects 2 voices for each combination of male and female and US and UK accents, 8 voices in total per model. Model creators "may also request" that specific voices be used where many are available, according to the same page. That is a meaningful lever for a vendor with a large voice library.

How is the arena scored?

Ratings come from a Bradley-Terry model fit by maximum likelihood to pairwise votes, anchored by fixing one model per arena at 1000. Votes are a binary choice with no tie option, model names are hidden, presentation order is randomised, and clips are loudness-normalised. Artificial Analysis publishes 95% confidence intervals and rank ranges, and most votes come from a paid panel screened for native English speakers resident in the US or UK.

The useful consequence: a one-point Elo gap means roughly a 50.1% chance that a listener picks the higher model (my arithmetic on the standard Elo scale). Against that, a 34-point gap to the third-placed Qwen-Audio-3.1-TTS-Plus (1296) implies about 55%. The top two are clearly above the field, but the field is closer than a "#1" label suggests.

What is the current leaderboard?

Captured 16 October 2026, all English (US and UK), all independent, all from Artificial Analysis. Prices are Artificial Analysis's per-1M-character figures as of that date.

The two Elo columns are separate fits on different arenas with different anchors. Do not subtract one from the other.

Is Eleven v4 still #1 on voice cloning and pronunciation?

No on both, as of 16 October 2026. The Controlled Voice Arena has every model clone the same 8 professionally recorded reference voices, which removes the confound of one vendor having nicer stock voices. There, HeyGen Voice (1198) and Qwen-Audio-3.1-TTS-Plus (1186) sit at rank 1-2, and the two ElevenLabs models share rank 3-5. At launch, Artificial Analysis had Eleven v4 at #2 on Controlled Voice with 1,157, so a later entrant moved it down.

The pronunciation benchmark shows the same drift. At launch, Artificial Analysis reported Eleven v4 at 91.7% (Eleven v3: 85.6%). Artificial Analysis later reported Fish Audio Drama 3 Preview at 93.2%, ahead of Eleven v4 at 91.7% and Eleven v4 Turbo at 90.1%.

The benchmark itself is narrow by design. Per Artificial Analysis, it uses 454 sentences containing 701 target words or phrases in four categories, read by a single pinned US English voice and judged by human reviewers. It tests whether "St." becomes "Saint" or "Street" correctly, not whether the delivery is expressive.

Where do ElevenLabs' numbers and the independent ones disagree?

ElevenLabs' preference and latency figures are company-reported and use comparator sets it chose. The launch post says Eleven v4 was preferred by roughly 75% of listeners, winning 65% to 81% of blind head-to-head tests, against Cartesia Sonic 3.6, Inworld TTS-2, Gemini 3.8 Flash-Lite TTS and Gemini 3.8 Flash TTS. Graders judged both expressiveness and naturalness, and ties counted as half. The footnote gives no listener count and no confidence intervals.

Two things stand out. First, the comparator list omits Qwen-Audio-3.1-TTS-Plus, which Artificial Analysis ranks third on Provider Voice. Second, the Elo gaps on Artificial Analysis imply smaller margins. By my arithmetic, a 44 to 82 point lead over Sonic 3.6, Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS translates to roughly 56% to 62% win probability, below ElevenLabs' 65% to 81%. The criteria, prompts and voices differ, so this is not a contradiction. It does show the two numbers should not be quoted as if they were the same test. (For more on this pattern, see how lab and independent scores diverge on OSWorld.)

Latency: 100ms or 150ms?

ElevenLabs gives two Turbo latency numbers. The models page says median inference latency is about 100ms, excluding application and network latency. The launch post says median time to first speech is about 150ms, measured in September 2026 with identical scripts and default settings, network latency measured and removed, over WebSocket streaming. The post's chart puts the other tested systems at 262 to 814ms.

Both are lab-reported. The Artificial Analysis methodology covers throughput, meaning characters of input converted per second, and does not report time to first audio. The models page lists no latency figure for Eleven v4.

Eleven v4 vs v4 Turbo: which should you use?

On the evidence available, Turbo is the better default unless you have a specific reason to pay for v4. It lists at half the price with a statistically indistinguishable arena score. Same-technology positioning comes from ElevenLabs: the company says Turbo "brings that same technology to low-latency use cases".

Pricing and limits come from the pricing page and models page. A 10,000-character script costs $0.80 on v4 and $0.40 on Turbo at list price, from the per-1K-character rates.

Migration notes from the v4 docs: Style and Speed sliders are not available, SSML is not supported, and when the target language differs from the reference voice's language, v4 generates native-sounding speech instead of carrying over the original accent. ElevenLabs also says audio tags are "not perfect yet".

What are the weaknesses of this ranking?

Five limits matter for anyone using it to buy or build.

• Narrow voice sample. Provider Voice uses 8 voices per model, US and UK English only, and creators may request which voices are used.
• Short-form bias is unstated, not ruled out. The methodology does not describe long-form tests up to the 10,000-character limit, which is where v4 is positioned for audiobooks.
• Preference is not accuracy. Elo captures what listeners prefer. The pronunciation benchmark captures one error type and shows a different leader.
• Language coverage. ElevenLabs claims 90+ languages. The boards cited here are English (US and UK), so the claim is untested in this analysis.
• Moving targets. Arena scores shift as votes accumulate and new entrants appear, which is why every figure here carries a capture date. For the general problem, see AI benchmark saturation in 2026.

Artificial Analysis states that no compensation is received from any providers for listing or favourable outcomes. Votes come from a paid panel, and the vendor-requested-voices option is the main place a conflict could enter.

Should you trust the Artificial Analysis TTS arena, and for what?

Trust it as a relative preference signal on English stock voices, not as a verdict on cloning, pronunciation, latency or multilingual quality. These scores are my assessment from the published methodology, not Artificial Analysis's own.

What would a better TTS comparison look like?

It would report the same model on long-form audio, multilingual prompts and cloned voices in one place, with latency measured independently. That means time to first audio at p50 and p95 under stated network conditions, a larger voice set, and ties retained in the vote data. Eleven v4's selling points are expressive control and 90+ languages, and neither is what a US and UK English preference arena is built to test. For why harness and setup choices change scores in other domains, see coding agent benchmarks and the harness effect.

What's still unclear?

• Which variant is better, if either. A one-point gap with overlapping intervals cannot separate them. ElevenLabs publishes no figure for what expressiveness Turbo gives up for speed.
• Independent latency. Neither the ~100ms nor the ~150ms figure has been reproduced by a third party.
• ElevenLabs' preference test. Listener count, confidence intervals and the reason Qwen-Audio-3.1-TTS-Plus was left out are not stated.
• Long-form and non-English standing. The English boards cannot confirm the multilingual and audiobook positioning.
• Snapshot drift. Scores quoted in the launch post, third-party explainers and the live board differ by up to about 13 points for the same model, and will keep moving.