ElevenLabs releases Eleven v4 speech models; v4 leads one Artificial Analysis arena and sits second on another

ElevenLabs released two text-to-speech models, Eleven v4 and its low-latency variant Eleven v4 Turbo, in a blog post published on 28 September 2026 and updated on 29 September. The company calls Eleven v4 "our most emotive text-to-speech model yet".

At a glance

• Both models "support more than 90 languages", according to ElevenLabs.
• ElevenLabs says Eleven v4 Turbo has "a median inference latency of ~100ms" and "a median time to first speech of ~150ms".
• ElevenLabs says Instant Voice Clones "can now capture voices with high fidelity using just 10 seconds of audio".
• On Artificial Analysis's Provider Voice leaderboard, Eleven v4 ranked first of 92 models when we read it on 30 September. On the same site's Controlled Voice leaderboard, which uses cloned voices, it ranked second of 38.

What ElevenLabs says is new

ElevenLabs says Eleven v4 is "Built on an entirely new architecture" and "was designed to interpret tone, pacing, emotion, character, and context". Users can steer delivery with inline tags such as "[laughs]" or "[light rain]", and the company says v4 "follows these audio tags and direction prompts more accurately than prior models".

Its developer documentation is more guarded on tags: "They're not perfect yet, and we're continuing to iterate and improve how reliably the model follows tag instructions". The same page says "Style and Speed sliders are not available in Eleven v4, and SSML is not supported", and that the model's "behavior may shift over time" as training continues.

The models page lists a 10,000 character limit for Eleven v4. The documentation says Eleven v4 is available through the Text to Dialogue API and Eleven v4 Turbo through the Text to Dialogue WebSocket.

Speed: two different medians

ElevenLabs gives two latency figures for Eleven v4 Turbo, and they measure different things:

• ~100ms median inference latency. The models page marks this figure "Excluding application & network latency".
• ~150ms median time to first speech. The blog's footnote defines this as "Median time from request to audible speech", measured in September 2026 against Cartesia Sonic 3.6, xAI TTS, Google Gemini 3.8 Flash-Lite TTS and OpenAI GPT-4o mini TTS, with "network latency measured and removed for all systems". The footnote does not print the rivals' figures.

Languages

The language list on ElevenLabs' models page for the v4 family names 91 languages, Welsh among them. TechCrunch reports that the previous version supported 70 languages and that the company "observed the biggest quality jump in Japanese, Brazilian Portuguese, Mandarin, and Cantonese".

The documentation describes a deliberate change: when a cloned voice speaks a different language from its source, v4 "generates fluent, natural-sounding speech in the target language — rather than carrying over the reference voice's accent". ElevenLabs says it is "exploring ways to make it a toggle".

Voice cloning

The blog's 10-second figure sits beside different guidance in the documentation, which says of Instant Voice Cloning: "Upload a short sample, generally one to two minutes". Eleven v4 also supports Professional Voice Clones.

ElevenLabs' launch post does not describe safeguards specific to v4. Its standing safety page says it blocks "the cloning of celebrity and other high risk voices" and requires "technological verification for access to our Professional Voice Cloning tool". Its documentation says "Voice-captcha technology is used to verify that Professional Voice Clones are created from your own voice samples."

The rankings, in full

ElevenLabs says v4 is "Ranked #1 by Artificial Analysis", citing the Provider Voice Arena Leaderboard, and "preferred by ~75% of listeners" in its own blind tests against four named models.

Artificial Analysis runs more than one leaderboard. The top five on each, as we read them on 30 September (Elo; price per 1M characters as listed by Artificial Analysis):

Provider Voice (each provider's own voices, 92 models)

1. ElevenLabs Eleven v4: 1316, $80.0
2. Cartesia Sonic 3.6: 1275, $49.0
3. Google Gemini 3.8 Flash TTS: 1268, $16.5
4. Alibaba Qwen-Audio-3.0-TTS-Plus: 1258, $19.3
5. Inworld Realtime TTS-2: 1247, $20.8

Controlled Voice (the same 8 cloned voices, 4 US and 4 UK; 38 models)

1. Alibaba Qwen-Audio-3.1-TTS-Plus: 1184, $19.3
2. ElevenLabs Eleven v4: 1154, $80.0
3. Inworld Realtime TTS-2: 1143, $20.8
4. Cartesia Sonic 3.6: 1139, $49.0
5. Alibaba Qwen-Audio-3.0-TTS-Plus: 1126, $19.3

Eleven v4 Turbo did not appear on either board when we read them. Arena rankings are built from ongoing votes and can change.

Price and availability

ElevenLabs says both models "are available now in ElevenAgents, ElevenCreative, and via ElevenAPI". Its pricing page offers a launch promotion: "On Creator plans and above, up to 2x your monthly TTS credits used on v4 won't count against your balance. Available in the web and mobile apps only." A site banner describes this as "3x credits included on Creator+ until October 12". We did not find a separate v4 per-character rate on ElevenLabs' pricing page.

Why it matters

On our reading, the release puts ElevenLabs at the top of Artificial Analysis's provider-voice arena, but not its cloned-voice arena, where an Alibaba model leads, and several close rivals are listed there at lower prices per character. ElevenLabs' preference and latency comparisons are its own tests, and the documentation's own caveats on tags and changing behaviour suggest testing before switching production voices.