Which ElevenLabs Models Should You Use? v3, v4, Multilingual and Flash

ElevenLabs models split into three jobs: expressive narration (Eleven v4 and Eleven v3), stable long-form speech (Multilingual v2) and fast, cheaper real-time output (Flash). Each one differs in language coverage, latency, character limit and price per character. This guide compares every current text to speech model, flags the deprecated ones, and matches each model to a concrete use case.

Last updated: October 2026. Plans and credit costs checked against ElevenLabs’ official pricing page.

What are ElevenLabs models?

ElevenLabs models are the speech engines that turn text, audio or prompts into output, and each product runs on its own family. Text to speech uses the Eleven, Multilingual and Flash models, transcription uses Scribe, and music uses the Music models. Picking one decides quality, speed, language reach and cost.

“Speech synthesis is the artificial production of human speech.”

Source: Wikipedia, Speech synthesis

Every model trades something. The most expressive models cost more per character and respond more slowly, while the fastest models are cheaper but less nuanced. The ElevenLabs review covers the platform as a whole; this page is only about choosing the right engine.

“Models control the quality, latency, and language coverage of generated audio.”

Source: ElevenLabs documentation

In short, a model is a trade-off between expressiveness, speed, languages and price.

ElevenLabs models at a glance

The current ElevenLabs text to speech lineup is six models: Eleven v4, Eleven v4 Turbo, Eleven v3, Eleven v3 Conversational, Multilingual v2 and the Flash pair. Two older Turbo models are deprecated. The table below lists each model’s language count, listed latency and character limit per request from the models documentation.

Source: ElevenLabs models documentation (elevenlabs.io/docs/models), checked October 2026.

Language reach and expressiveness rise with the newer Eleven models, while Flash wins on speed and on how much text fits into one request.

Five factors separate the models in practice:

• Expressiveness: Eleven v4 is described as the most emotionally rich model, while Flash prioritizes speed, so emotional range drops as response time gets shorter.
• Language coverage: Counts range from English-only Flash v2 up to 90+ languages on Eleven v4, so the target audience can rule models in or out.
• Latency: Real-time models list response times between roughly 75ms and 280ms, which matters for voice agents but rarely for recorded voiceovers.
• Character limit: Requests range from 5,000 characters on Eleven v3 to 40,000 on Flash v2.5, which decides how often long scripts must be split.
• Price per character: API prices run from $0.04 to $0.08 per 1,000 characters, so the faster models cost about half as much per generation.

Start from the use case, then use these five factors to narrow the list to one or two models.

What is Eleven v4 best for?

Eleven v4 is ElevenLabs’ newest flagship text to speech model, described in the docs as its most emotionally rich, expressive speech synthesis model. It covers 90+ languages, accepts up to 10,000 characters per request and supports audio tags, which makes it the default pick for narration where performance matters most.

The best uses are audiobooks, story-driven YouTube videos, character dialogue and ads where delivery sells the message. Audio tags such as whispers or laughs let a script direct emotion line by line, which the ElevenLabs text to speech settings guide covers in detail. On the API it costs $0.08 per 1,000 characters, the top of the price range.

Its trade-offs are cost and the lack of SSML break tags, so pauses come from punctuation instead.

Choose Eleven v4 when expressiveness matters more than speed or price.

Eleven v4 Turbo and v3 Conversational for real-time use

Eleven v4 Turbo and Eleven v3 Conversational are real-time versions of the expressive models, built for live conversations rather than recorded files. The docs list about 100ms latency for v4 Turbo and about 280ms for v3 Conversational, and both cost $0.04 per 1,000 characters on the API, half the price of their full models.

These models suit voice agents, interactive apps and games, where a reply has to start quickly but still sound expressive. That makes them natural partners for the ElevenLabs Agents platform, which bills by conversation minute. Character limits for both are not stated in the docs, so long single requests are better sent to the full models.

For live use, v4 Turbo offers the better listed latency of the two expressive real-time options.

When does Eleven v3 still make sense?

Eleven v3 is the previous expressive model, described in the docs as human-like and expressive speech generation across 70+ languages. It still makes sense when a project depends on its native IPA pronunciation support or when an existing workflow was tuned around its voices. Its 5,000-character limit is the tightest in the lineup.

ElevenLabs’ docs note that v3 does not offer the speed, similarity or Speaker Boost controls, so tuning relies on stability, audio tags and script writing. It costs the same $0.08 per 1,000 characters as v4 on the API, so for new projects with no special pronunciation needs, v4 is usually the more logical choice.

Keep v3 for IPA-heavy scripts and established workflows; start new projects on v4.

Is Multilingual v2 still worth using?

Multilingual v2 is a previous-generation model that the docs describe as lifelike with rich emotional expression, and it is not deprecated. It supports 29 languages and 10,000 characters per request. It remains worth using for long, steady reads where consistency between generations matters more than dramatic range.

Two practical advantages keep it relevant. It supports break tags for exact pauses, which v4 and v3 do not, and the pricing page lists a clear rate of about one credit per character in the web app. Some voices also simply pair better with it, which the docs say is true of every model.

It does not use audio tags, and its 29-language coverage is narrower than the Eleven models.

Multilingual v2 is the safe choice for consistent long-form narration in its supported languages.

Flash v2.5 and Flash v2 for speed and volume

Flash v2.5 and Flash v2 are ElevenLabs’ ultra-fast models, optimized for real-time use with about 75ms of listed latency. Flash v2.5 covers 32 languages and accepts 40,000 characters per request, while Flash v2 is English only with a 30,000-character limit. Both cost $0.04 per 1,000 characters on the API.

Flash suits high-volume work where cost and speed matter more than nuance: product narration at scale, notifications, voice agents and prototypes. In the web app the pricing page lists Flash and Turbo at between half a credit and one credit per character. Developers wiring it into code can follow the ElevenLabs API quickstart.

Flash is the cheapest and fastest way to generate a lot of speech.

Which ElevenLabs models are deprecated?

Deprecated models are older engines ElevenLabs has marked for replacement, and new projects should avoid them. The models documentation lists both legacy Turbo text to speech models as deprecated, along with the first Scribe transcription model, and describes the first music model as outclassed by its successors.

Five older models now have a recommended replacement:

• Turbo v2.5: Deprecated in the models documentation, with Flash v2.5 named as the replacement for multilingual real-time speech across the same 32 languages.
• Turbo v2: Deprecated English-only real-time model, replaced by Flash v2, which offers the same English focus and listed latency of about 75ms.
• Scribe v1: Deprecated speech recognition model, replaced by Scribe v2, which keeps 90+ language coverage and adds realtime and medical variants.
• Music v1: Described as outclassed by Music v2 and Music v2.5, so new music generation should start on one of the newer models.
• Multilingual v2: Not deprecated but labeled previous generation, which signals that newer Eleven models now carry most of the platform’s development focus.

Migrating from a deprecated Turbo model is usually a straight swap to the matching Flash model.

How many languages does each model support?

Language support is the number of languages a model can speak, and it varies sharply across the lineup. Eleven v4 and v4 Turbo list 90+ languages, Eleven v3 lists 70+, Flash v2.5 lists 32, Multilingual v2 lists 29, and Flash v2 handles English only. That is why a single language count for ElevenLabs is misleading.

Different official pages quote different totals, so the safest rule is to always tie a language count to a named model. Four checks help match a model to an audience:

• Target language list: Confirming the exact language on the model’s documentation entry prevents relying on a platform-wide headline number that may belong to another model.
• English-only scripts: ElevenLabs’ docs strongly recommend an English-only model for exclusively English output, which points to Flash v2 for speed or a full model for quality.
• Mixed-language scripts: Multilingual or Eleven models handle scripts that switch languages, while English-only models will mispronounce or skip the non-English passages.
• Language code hint: The API’s optional language code parameter helps text normalization apply the right rules for short or ambiguous text in a given language.
• Voice and accent fit: A voice recorded in one language may carry its accent into others, so testing the chosen voice in each target language avoids surprises.

Name the model whenever a language count is quoted, and test the voice in every target language.

Credit and API cost per model

Cost per model is the price of generating speech, charged in credits in the web app and in dollars per 1,000 characters on the API. Per the ElevenLabs API pricing page, the full expressive models cost $0.08 per 1,000 characters, while the Turbo, Conversational and Flash models cost $0.04.

Source: ElevenLabs API pricing (elevenlabs.io/pricing/api) and pricing page (elevenlabs.io/pricing), checked October 2026.

Switching from a full model to its fast counterpart halves the API bill for the same text.

Web app users pay in credits instead. Using the pricing page’s estimate of about 10 minutes of speech per 10,000 credits, a 10-minute voiceover on Multilingual v2 uses roughly 10,000 credits. Plan-by-plan totals are in the ElevenLabs pricing breakdown.

Fast models cost about half as much, so reserve the full models for audio where expression earns its price.

Speech to text and music models

Beyond text to speech, ElevenLabs runs separate model families for transcription and music. Scribe v2 handles speech to text across 90+ languages, with realtime and medical variants, while Music v2.5 and Music v2 generate songs from prompts. Each family has its own credit and API pricing.

Five non-TTS models are worth knowing:

• Scribe v2: The standard speech recognition model covering 90+ languages, priced at $0.22 per hour on the API and about 330 credits per minute in the app.
• Scribe v2 Realtime: A real-time transcription model with about 150ms listed latency, priced at $0.39 per hour on the API for live captioning and agent use.
• Scribe v2 Medical: A recognition model specialized for medical and clinical audio, sharing the $0.22 per hour API price with the standard Scribe v2 model.
• Music v2.5: Described as ElevenLabs’ most advanced music model, generating tracks from text prompts at about 900 credits per minute in the app.
• Music v2: Studio-grade music generation from text prompts, composition plans and previously generated songs, priced at $0.15 per minute on the API.

ElevenLabs’ homepage claims 98% accuracy for Scribe without stating test conditions, so treat it as a vendor claim. Details live in the ElevenLabs speech to text guide and the ElevenLabs Music guide.

Transcription and music each have dedicated models, so text to speech choices do not carry over.

Choosing ElevenLabs models by use case

Choosing by use case is the fastest way to pick a model, because each project type rewards one trade-off over the others. Recorded narration rewards expressiveness, live agents reward latency, and bulk generation rewards a low price per character. Matching the job first avoids paying for quality nobody hears.

Six common projects map cleanly to a model:

• Audiobooks and long narration: Eleven v4 for emotional range, or Multilingual v2 when steady consistency across many chapters matters more than dramatic delivery.
• YouTube voiceovers: Eleven v4 for story-driven channels, while explainer and tutorial channels can save credits with Multilingual v2 or Flash v2.5.
• Voice agents and phone bots: Eleven v4 Turbo for expressive real-time replies, or Flash v2.5 when the lowest listed latency and cost come first.
• Multilingual localization: Eleven v4 for its 90+ language coverage, checking each target language on the models page before committing to a full project.
• High-volume product audio: Flash v2.5 at $0.04 per 1,000 characters, with its 40,000-character limit reducing the number of separate requests needed.
• Pronunciation-critical scripts: Eleven v3 for native IPA input, which gives exact control over brand names, medical terms and foreign words.

Pick the model that serves the listener’s experience, then check cost and limits.

Bottom line: which ElevenLabs models to use

The best ElevenLabs model for most creators is Eleven v4 for expressive recorded audio and Flash v2.5 for fast or high-volume work. Multilingual v2 stays useful for consistent long reads, Eleven v3 for IPA pronunciation, and the Turbo and Conversational models for live agents. Avoid the deprecated Turbo models entirely.

“Choosing an ElevenLabs model is a trade between expression, speed and language reach.”

Source: GrowthStackKit, ElevenLabs Review

Model choice also affects how far a plan’s credits stretch, which the guide to ElevenLabs credits explains with worked examples. Test the same paragraph on two candidate models before committing a whole project.

Start with v4 or Flash, then switch only when a specific need points elsewhere.

Frequently asked questions

The 12 most-asked questions about ElevenLabs models.

What is the newest ElevenLabs model?

Eleven v4 is the newest flagship text to speech model in the documentation, described as ElevenLabs’ most emotionally rich, expressive speech synthesis model. It covers 90+ languages and accepts 10,000 characters per request. Eleven v4 Turbo is its real-time variant, with about 100ms of listed latency.

Which ElevenLabs model is the fastest?

The Flash models are the fastest, with about 75ms of listed latency according to the models documentation. Flash v2.5 covers 32 languages and Flash v2 is English only. The listed figure excludes application and network latency, so real response times in an app will be longer.

Which ElevenLabs model is the cheapest?

Flash, Turbo, Eleven v4 Turbo and Eleven v3 Conversational cost $0.04 per 1,000 characters on the API, half the $0.08 rate of Eleven v4, Eleven v3 and Multilingual v2. In the web app, Flash and Turbo use between half a credit and one credit per character.

Is Multilingual v2 deprecated?

No. The models documentation labels Multilingual v2 as a previous-generation model, but it is not deprecated. It supports 29 languages and 10,000 characters per request, and it remains a solid choice for consistent long-form narration and scripts that rely on break tags for pauses.

How many languages does ElevenLabs support?

It depends on the model. Eleven v4 and v4 Turbo list 90+ languages, Eleven v3 lists 70+, Flash v2.5 lists 32, Multilingual v2 lists 29, and Flash v2 is English only. Always check the specific model rather than relying on one platform-wide number.

Should I use Eleven v3 or Eleven v4?

Eleven v4 is the better default for new projects because it is newer, covers more languages and allows longer requests at the same $0.08 API price. Eleven v3 still makes sense for scripts that depend on its native IPA pronunciation support or for workflows already tuned around it.

Which model is best for voice agents?

Real-time models fit voice agents best. Eleven v4 Turbo offers expressive speech with about 100ms of listed latency, Eleven v3 Conversational lists about 280ms, and Flash v2.5 lists about 75ms. All three cost $0.04 per 1,000 characters on the API.

What replaced the ElevenLabs Turbo models?

The models documentation marks Turbo v2.5 and Turbo v2 as deprecated. Flash v2.5 replaces Turbo v2.5 for multilingual real-time speech, and Flash v2 replaces Turbo v2 for English-only use. The swap usually needs no other workflow changes.

Do all models support audio tags?

No. According to ElevenLabs’ docs, audio tags work on Eleven v4, Eleven v4 Turbo and Eleven v3. Multilingual v2 does not use audio tags, and relies on break tags and punctuation for pacing instead. Flash models are tuned for speed rather than tag-driven delivery.

What is the character limit per request?

The models documentation lists 10,000 characters for Eleven v4 and Multilingual v2, 5,000 for Eleven v3, 40,000 for Flash v2.5 and 30,000 for Flash v2. Limits for Eleven v4 Turbo and Eleven v3 Conversational are not stated, so long scripts are best split.

Which model does ElevenLabs use for transcription?

Scribe v2 is the current speech recognition model, covering 90+ languages at $0.22 per hour on the API. Scribe v2 Realtime handles live transcription with about 150ms of listed latency, and Scribe v2 Medical targets clinical audio. Scribe v1 is deprecated.

Does the model change the cost in credits?

Yes. The pricing page lists Multilingual v2 at about one credit per character and Flash and Turbo at between half a credit and one credit per character. On the API, the full expressive models cost twice as much per 1,000 characters as their fast counterparts.