Eleven v4

Initial reactions arrived on Japanese, Turkish, Spanish, and German the evening and night after release. They praise the emotions, clones, and other languages, while criticizing the price and noise on screams and breathing.

Eleven v4 is available even on the free ElevenLabs plan. You cannot download it; the model runs on the website, in the vendor's apps, and via API.

You create videos, courses, or books and want natural intonation. In v4, tags such as [whispers] or [laughs] set it directly in the text.

You are building a voice bot or assistant. According to ElevenLabs' measurement, Turbo starts speaking after 0.15 seconds.

You want to generate speech in another language using your own voice. A v4 clone speaks it without the accent of your original recording.

You generate a large volume of text without emotions. After Oct 12, v4 costs twice as much as Flash v2.5.

Getting started is covered in chapter 4. The choice between v4, Turbo, and Flash is covered in chapter 8.

On the first day, practitioners used v4 for videos, games, and clones of their own voices. Every case below includes audio: play it in the player.

A video in half an hour. Japanese creator Nakazaki Kobo gave the API key for Opus 5.5 to Claude Code and asked for a Japanese v4 voice-over test video. The finished video arrived 30 minutes later, virtually on the first try.

A call that switches to Chinese. In ElevenLabs' own demo, an agent powered by Turbo handles a call about medication approval and, at the caller's request, switches from English to Chinese in the same voice. ElevenLabs measures Turbo starting to speak in 0.15 seconds; OpenAI GPT-4o mini TTS takes 0.81.

Where it stumbled. Toshiya Nakamura, who runs corporate AI courses, had v4 perform emotions in a voice actor's voice. There were so many emotions that the original voice's character fell apart, and the author could not decide whether that was bad.

More. Tatiana Tsiguleva, creative director at Restream, tested a clone of her voice, and it spoke without her accent. The developer of indie game Second Dawn gave its trailer a v4 narrator on release night, while meme creator Ezra voiced an almost six-minute video entirely with v4 on the free plan.

v4 launched simultaneously on the ElevenLabs website, in apps, and in the API; Turbo launched in the API and voice agents. The model IDs in code are eleven_v4 and eleven_v4_turbo, and switching models takes one line.

Free plan. The vendor’s anti-abuse protection may disable Free if another account has already logged in from the same address. Paid plans are not affected by this protection.

Older clones. The vendor recommends retraining voice clones made before v4 on the new model. Otherwise, they sound worse.

Pauses and markup. SSML such as <break> is disabled in v4; use [pause] or [long pause] for a pause. Pronunciation dictionaries for names and terms continue to work.

In Artificial Analysis' blind listener vote, v4 took first place out of 92 models. It compared vendors' own English voices: listeners choose which recording sounds more natural.

In the ranking where all models read in the same cloned voices, Alibaba's Qwen-Audio 3.1 is first: 1182 versus 1159 for v4. It costs about $19 per million characters, one-quarter the price.

Turbo is not in Artificial Analysis' rankings. Its 0.15 seconds to first sound is ElevenLabs' own measurement; Speko, a gateway that sells access to the model, measured 205 milliseconds.

In the API, v4 costs $0.08 per 1,000 characters, the same as v3, while Turbo costs half as much. A 72% discount is available through Oct 12: $0.022 and $0.011.

Subscription. The website uses credits: Free gives 10,000 per month, around 10 minutes of v4 speech. Starter for $6 unlocks voice cloning and commercial use.

Promotion. On plans from Creator at $22, v4 in the apps does not use credits for two weeks, up to twice the monthly allowance.

Under ElevenLabs rules, every v4 clone requires verified consent from the voice owner. The vendor says its detector recognizes such recordings, and each one can be traced back to its creator.

Data. Texts and tags sent to v4 are not used for model training without the customer’s consent. A zero-data-retention mode is available only to enterprise customers.

Limitations. v4 has no style or speed sliders, and in another language a voice always speaks without the accent of the original recording. The vendor warns that model behavior may change.

Sound. Users report a hiss beneath the voice and artifacts in screams and heavy breathing. Overacting also occurs; see an example in chapter 3.

Within ElevenLabs, the choice is simple: v4 for recorded videos and books, Turbo for live conversation. Flash v2.5 is faster than both and costs the same as Turbo, but ranks 45th with listeners.

With Gemini. Japanese practitioners compared v4 with Gemini 3.8 Flash TTS on the first day. Gemini wins on price; v4 on emotions, clones, and accent.

With v3. The vendor explicitly recommends switching to v4. The API price is the same, while in our test v4 generated the full file more than twice as fast.

v4 arrived eight months after v3 left alpha. In three years, the lineup grew from 29 languages to more than 90.