Generative Footage

ElevenLabs Voice Model Tiers for Video Voiceover

ElevenLabs now offers five models that trade speed, expressiveness, and language coverage.

· 8 min read · Updated
Cover illustration for “ElevenLabs Voice Model Tiers for Video Voiceover”
Model Comparisons · September 18, 2026 · 8 min read · 1,826 words

ElevenLabs is no longer a single tool that turns text into speech. It now runs as a stack of tiers, and the tier picked at the start of a project sets the latency, the expressiveness, the language range, and the cost of everything that follows. That choice has to happen before a voice gets picked, before a script gets formatted, before any other step in production. The platform behind this decision has grown wide: text-to-speech sits alongside speech-to-text, voice cloning, dubbing, agents, a Studio editor, sound effects, and music generation, and a stack that wide can't run on one default model anymore. A TikTok voiceover, a dubbed training module, and a real-time customer support agent each need a different tier to do their job well, and treating them the same costs either quality, money, or both.

The current model lineup and the three axes that separate the tiers

Three things separate the tiers from each other: how fast a model responds, how expressive it sounds, and how many languages it covers. Each tier represents a trade a creator is making. The current production stack reflects ElevenLabs' September 28, 2026 launch and official documentation.

Eleven v4 is at the top of the quality tier. It's the most emotive model in the lineup, covers more than 90 languages, allows a high character count per request, supports multi-speaker dialogue, and now stands as the flagship model for quality work. The fast version of that same model, Eleven v4 Turbo, is built for real-time use with a median inference latency around 100 milliseconds and the same 90-plus language coverage. Eleven v3 still delivers emotionally rich, dramatic reads with support for audio tags like [laughs], [whispers], and [pause], covers more than 70 languages, and handles multi-speaker dialogue, but it has moved down to second place in the quality tier now that v4 has taken the top spot. Eleven v3 Conversational trades some of that richness for speed, running at roughly 280 milliseconds of latency while keeping the same 70-plus languages and audio tag support. Eleven Multilingual v2 rounds out the stack as the most stable option for long narration, lifelike and consistent in quality, though its language coverage tops out at 29.

The three axes work like this: on speed, Flash v2.5 and v4 Turbo run in real time, while v3 and Multilingual v2, both running at roughly a second or more, only work for offline rendering. On expressiveness, v4 leads the pack, v3 follows close behind with its audio tag support, Multilingual v2 sounds lifelike but emotionally flatter, and Flash v2.5 gives up expressiveness in exchange for its speed. On language reach, v4 and v4 Turbo cover more than 90 languages, v3 covers more than 70, Multilingual v2 covers 29, and Flash v2.5 covers 32. A fourth factor appears only in the billing, not on the feature list, but it matters just as much at scale: Flash models cost only 0.5 credits per character, which in practice doubles a creator's monthly output on the same plan compared to standard models.

Diagram: Speed, Expressiveness, and Language Coverage Across the Model Tier Stack. Visualizes: Visualize how five ElevenLabs models trade off against three axes: speed (latency), expressiveness, and language coverage.

How Eleven v4's arrival at the top of the quality tier changes the decision for high-stakes voiceover

Eleven v4 launched on September 28, 2026, and its arrival pushes v3 out of the top spot in the quality tier, which raises the bar for what ads, branded narration, and other long-form, high-stakes voiceover work can sound like when it comes out of an AI pipeline. The launch came with an early signal of just how strong the model is: on Artificial Analysis's Speech Arena, Eleven v4 ranked first with an Elo score of 1319, ahead of Cartesia Sonic 3.6 at 1276 and Gemini 3.8 Flash TTS at 1267. That's the first independent benchmark reading on the model, recorded at launch.

ElevenLabs describes V4 Turbo as "the low latency version of Eleven v4," meaning the expressiveness gains of v4 are now partially accessible in the near-real-time tier, including offline rendering. For video creators doing premium narration work, the decision that used to point to v3 before September 28, 2026 now points to v4 instead, with v3 staying useful only as a fallback for integrations where v4 hasn't rolled out yet.

That said, the benchmark is a launch-day reading, not a long production history. Nobody has run independent long-form assessments of v4 yet, so creators handling extended narration work should test the model against their own material before committing a full project to it.

Short-form social video: why the quality tier is often the wrong choice

For TikTok, Reels, and YouTube Shorts, the best-sounding model is often the wrong pick. Short runtimes, the attention economics of the format, and the sheer volume creators need to produce make Flash v2.5, or v3 run at lower stability settings, the smarter operational choice over v4, even though v4 is the stronger model on paper. Short-form viewers are used to hearing delivery shift and move.

A creator posting shorts daily will burn through a Creator plan's monthly credits fast if every clip runs on a standard-tier model, but Flash v2.5's rate of 0.5 credits per character effectively doubles that capacity without needing to upgrade the plan. Flash v2.5 also carries a 40,000-character limit, eight times what v3 allows. A creator can generate a full week's worth of scripts in one single API call, something v3 simply can't do. For short-form work specifically, running v3 at a slightly lower stability setting than would be used for long-form narration produces the delivery variation that short-form audiences respond to. The delivery variation comes from a settings adjustment at the same model tier.

Tools that let a creator plug ElevenLabs into a broader video workflow make this easier to manage at scale. Set the model and voice parameters once to generate audio across an entire batch of short-form content without switching between separate applications.

Training videos and e-learning modules: where Multilingual v2 still earns its place

Training videos and e-learning content call for something different than either speed or peak expressiveness. Eleven Multilingual v2 remains the strongest default for this format because its long-form stability, consistent sound across a full course, and solid language coverage keep narration steady over a full module. Most training modules share the same shape: one narrator, a scripted delivery, a steady pace from start to finish, and that's exactly the pattern Multilingual v2 was built for, backed by a 10,000-character limit and a track record of staying stable over long stretches of narration.

That stability pays off especially well in localization. A company can take one training video, run it through Multilingual v2's dubbing workflow, and roll it out across offices in different regions while keeping the quality and the message consistent, all within the model's 29-language range. There's a practical advantage buried in how these projects get maintained too: when a module needs an update, regenerating a single paragraph with the same voice and the same settings costs far less than rebooking a studio session, as long as the production team kept a record of the voice name, model, and settings alongside the original script file.

Multilingual v2 may not hold this position forever. If independent testing eventually confirms that v4 holds up as well at sustained length, its coverage of more than 90 languages would make it the stronger pick for localization work. For now, v2 simply has the longer track record in this specific format. Teams producing training libraries at scale also get infrastructure that matters here: ElevenCreative offers multi-seat workspaces, shared credit pools, content libraries, and role-based approvals, though that tier of access only opens up at the Scale plan and above.

Ads and branded commercial reads: the case for v4 and Professional Voice Cloning

Ad and commercial voiceover is the one format where paying for the top tier clearly pays for itself. A voice that sounds even slightly synthetic in a 30-second spot costs a brand in conversion, not just in production time. Because ads run short, the added credit cost of using v4 barely registers compared to a long-form project, and there's no volume pressure pushing a creator toward Flash the way there is in short-form social content.

Generation settings also shape whether the read lands. For ad voiceover, Stability in the 35 to 50 range keeps the delivery natural without sounding too predictable, Similarity between 75 and 85 keeps the voice faithful to the original clone, and Style between 0 and 30 controls how much emotion comes through. These settings trade some consistency for expressiveness, which fits the short, punchy nature of commercial reads. Script prep matters just as much as the model and the settings: keeping sentences short is standard practice for ad copy, because long subordinate clauses break up intonation no matter which model is reading them.

Voice cloning adds another layer to this decision. Professional Voice Cloning, available starting at the Creator tier, produces a voice that's genuinely difficult to tell apart from the original speaker, and for branded work where a voice functions as part of the brand itself, PVC is the production-grade option over Instant Voice Cloning. The free plan doesn't allow commercial use at all, and the minimum plan that unlocks it, whether for monetized YouTube content, client work, or advertising, is the Starter plan at $6 a month. Platforms that connect ElevenLabs directly into a video editing workflow also cut down the back-and-forth of moving audio files between separate tools, which speeds up the path from a generated voiceover to a finished cut.

Dubbed explainers and multilingual content: when language coverage forces the tier decision

Multilingual dubbing and localized explainer content work differently from the other formats on this list: the deciding factor is which languages a model actually covers, and that alone determines which tiers are even on the table. v4 and v4 Turbo cover more than 90 languages, v3 covers more than 70, and Multilingual v2 covers 29.

Before v4 existed, a creator dubbing into a language outside Multilingual v2's 29-language range had to settle for Flash's lower expressiveness just to get coverage. With v4 now covering more than 90 languages at full quality-tier expressiveness, that trade-off disappears for most target languages. Choosing a voice model for multilingual work in 2026 isn't a one-vendor decision anymore. ElevenLabs leads on cloning quality and natural sound, but the right model still depends on the job at hand: an audiobook chapter and a multilingual dubbing pass call for different answers even within ElevenLabs' own lineup.

Dubbing workflows that handle translation and re-voicing as a built-in part of the platform let a team make the tier decision once, at the platform level, instead of re-deciding it for every file that comes through. One more constraint shapes this choice for dubbed content specifically: audio tags like [laughs] and [whispers] work in v3, v4, and v4 Turbo, but not in Multilingual v2. Any dubbed project that needs those expressive cues belongs in v3 or v4, regardless of which language it's being dubbed into.

Sources

  1. ElevenLabs Voice Cloning Review 2026: v3, Scribe & Agents | Coval
  2. Text to Speech | ElevenLabs Documentation
  3. ElevenLabs Pricing (September 2026): Compare Plans & Find the Best One
  4. Eleven v4: The New #1 Voice Model, Turbo and Price

More in Model Comparisons