Source: ElevenLabs
A new AI voice model has arrived that pauses for breath, laughs, and speaks with feeling, the way a person does. On September 28, 2026 at 23:01 KST, ElevenLabs released Eleven v4, a high-quality speech synthesis model, and Eleven v4 Turbo, a low-latency version, calling them "the fastest and most expressive voice models." They capture small nuances of the voice to sound like a recording of a real conversation, and they support more than 90 languages, including Korean. The official video is worth listening to from the 14-second mark to hear the difference.
Source: X @ElevenLabs
How Eleven v4 and v4 Turbo differ
Both models use a new architecture so tone, pace, emotion, and character come out as you intend. v4 prioritizes expressiveness and audio quality of the output, while Turbo cuts latency for real-time voice agents.
| Model | Best for | API model_id | Latency and features |
|---|---|---|---|
| Eleven v4 | Content production, narration, conversational voice | eleven_v4 | Up to 10,000 characters per request, streaming supported |
| Eleven v4 Turbo | Real-time support, sales, and booking agents | eleven_v4_turbo | Median inference latency of about 100ms, about 150ms to first audio, streaming supported |
Turbo suits real-time agents where conversational flow matters, like support or booking. For example, when you build an experience that goes back and forth with customers right away, like Revolut's ElevenLabs AI voice support, it is an option for cutting latency. If the lowest latency and price come first, you can also consider Flash v2.5 at about 75ms latency, but it supports 32 languages, fewer than the v4 family.
Source: X @ElevenLabs
In the Provider Voice TTS Arena ranking Artificial Analysis published on September 28, 2026, v4 took first place overall with an Elo of 1,319 and also ranked first in pronunciation robustness at 91.7%.
Source: X @ArtificialAnlys
Direct the performance with inline tags
Put tags like [laughs], [whispers], [door slams], or [said angrily in French accent] inside your script to specify delivery, emotion, pace, reactions, sound effects, and style. You can also add direction in natural language. It follows tag order and sound effects more precisely than v3, and IPA pronunciation control has improved too.
For example, a short script can be put together like this.
[whispers] Tell me the second the door opens.
[door slams] [gasps] Now! [laughs nervously]Source: X @ElevenLabs
Source: X @NevFlynn
Voice cloning in 10 seconds
Instant Voice Clone clones a voice from 10 seconds of audio. Professional Voice Clone suits cases that need higher fidelity. v4 has been improved to keep the same voice identity across long narration and conversation. Cloned voices made before v4 are best retrained to get good results on v4. The v4 voice library you can choose from has more than 17,500 voices.
Source: X @ElevenLabs
90+ languages, including Korean
Eleven v4 and Turbo support more than 90 languages, including Korean (ko). The official thread named Cantonese, Mongolian, and Odia as examples of supported languages. The previous generation, v3, supports more than 70 languages, and Flash v2.5 supports 32.
Source: X @ElevenLabs
Use it in the web app and the API
On the web, open ElevenLabs Text to Speech, pick your text and voice, then choose Eleven v4 as the model and generate.
In the API, set the request's model_id to eleven_v4 or eleven_v4_turbo. You can use the same ID in the Python and TypeScript SDKs, the ElevenLabs CLI, the MCP server, and ElevenAgents. A single v4 request is limited to 10,000 characters, so split longer text into chunks and pass previous_text and next_text to carry context across.
from elevenlabs import ElevenLabs
client = ElevenLabs()
audio = client.text_to_speech.convert(
text="Hello. This is a voice made with Eleven v4.",
voice_id="YOUR_VOICE_ID",
model_id="eleven_v4",
)
If you do not have the voice you want, you can first use Voice Design, which creates a voice from a text description. Voice Design is not new in v4 but a feature offered separately, and you can use the voices you create there for v4 generation.
Source: X @ElevenLabsDevs
Pricing and the launch discount
In the ElevenCreative web app, Creator and higher subscribers can use v4 free for up to 2 times their monthly credits. For the API, through October 12, 2026, the launch discount is $0.022 per 1,000 characters for v4 and $0.011 for Turbo. That comes to $22 and $11 per million characters, and list prices are $0.08 and $0.04 per 1,000 characters, respectively.
| Plan | Monthly price | Characters included per month |
|---|---|---|
| Free | $0 | 10,000 |
| Starter | $6 | 273,000 |
| Creator | $22 ($11 first month) | 1,000,000 |
| Pro | $99 | 4,500,000 |
| Scale | $299 | 13,591,000 |
| Business | $990 | 45,000,000 |
The monthly inclusions for the plans are as of the September 29, 2026 pricing page. This summarizes the characters on the currently published pricing table together with the v4 API discount price.
Related posts
You can look up terms that come up often, like AI models and speech synthesis, in the AI and automation glossary.