ElevenLabs v4 can act. Now your script has to direct it

In the small hours of Tuesday in Japan, less than four hours after ElevenLabs released its new voice model, the AI anime director k_matsumaru sat down with two characters from an original anime he had already made and gave them a scene to play. He typed the lines by hand, ran them through Eleven v4 once, and posted the result without an edit. By his own account the whole voice drama took about an hour from idea to upload. His verdict on X was blunt: "The acting is ridiculously good, and the voice fidelity is amazing too." He had tried Irodori-TTS and Gemini 3.8 TTS on the same kind of material, and v4 came out, in his words, streets ahead.

That post has drawn about 174,000 impressions and 1,300 bookmarks, which makes it the most useful piece of evidence on the internet right now for anyone asking the question that matters to filmmakers: can a synthetic voice finally carry a performance? The honest answer after a day of checking is mostly yes, with two footnotes that ElevenLabs' launch copy leaves out, and one consequence that changes where the work in an AI short actually happens.

What shipped on Monday

ElevenLabs launched two models on Monday 28 September at 14:01 UTC. Eleven v4 is the quality model for narration, dialogue and character work, taking up to 10,000 characters per request with several speakers in one pass. Eleven v4 Turbo is the fast one, built for live voice agents, with a median inference time of about 100 milliseconds according to the company. An independent benchmark from Speko measured Turbo at 205 milliseconds on its own test, which is still quick but not the headline number. Both models went live across the company's apps and its API (the interface developers build on) the same day, Runway added v4 that evening, and fal lists it too.

The languages page in the documentation lists 91 languages, up from 70 or more for v3, and we counted them. The creative change is direction. Inline audio tags such as [whispers], [laughs] or [said angrily in French accent] sit inside the script, alongside sound cues like [light rain], and lines can be steered in plain language. ElevenLabs says v4 follows those tags more accurately than earlier models. k_matsumaru made the same point from the other side, writing that v3 was not bad but that v4's fidelity to the reference voice and the freedom of its tags are on another level, and that it often gives a good read with no tags at all.

Number one, with a footnote

The launch leans on a ranking from Artificial Analysis, and the ranking is real. On its Provider Voice Arena, read on Tuesday, Eleven v4 sits first at 1,315 Elo, ahead of Cartesia Sonic 3.6 and Gemini 3.8 Flash TTS, while Eleven v3 sits 17th. That is a large jump over the company's own previous flagship. The footnote is that this board scores each provider on its own stock voices. On the Controlled Voice board, where every model reads with the same eight cloned voices, v4 is second at 1,159, behind Alibaba's Qwen-Audio-3.1 at 1,182, a model that costs $19.30 per million characters.

That second board is the one that matters if you plan to clone your own voice or an actor's. It is also worth knowing that the roughly 75% preference figure in the launch post comes from ElevenLabs' own blind test, which compared v4 with Cartesia, Inworld and two Gemini models and left Qwen out.

Ten seconds in the blog, a minute in the manual

The most repeated claim from launch day is that v4 can clone a voice from ten seconds of audio. It rests on a single sentence in the blog and the launch thread, and we found no independent test of a ten-second clone. The company's own Instant Voice Cloning guide, read on Tuesday, still tells users to record at least one minute and recommends one to two, so the documentation and the headline disagree on day one. Treat ten seconds as the floor the system accepts, not a promise about how the clone will sound.

The consent side has not moved either. An instant clone still asks the user to tick a box confirming they have the right and consent to clone the voice, a self-attested check, while professional clones are limited to your own voice and require a spoken verification. As the amount of audio needed falls, that checkbox carries more weight, and nothing announced on Monday adds to it.

What a voice track costs now

Pricing is confirmed. Until Monday 12 October the API costs $22 per million characters for v4 and $11 for Turbo, and after that the list price is $80 and $40. Subscribers on the Creator plan and above get v4 free in the company's apps for the same two weeks, up to twice their monthly credits. fal charges the list rate, so the discount only applies when you buy direct.

By our math, a ten-minute short at around 150 spoken words a minute is about 9,000 characters. One full pass costs about 20 cents at the launch price and 72 cents at list, and voicing the same film in ten languages comes to roughly $7.20 at list, for the speech alone. Translation, retakes and the time spent listening are not in that figure, and retakes are where the money goes. At list, v4 is premium for the category, around five times Gemini 3.8 Flash TTS and 1.6 times Cartesia, but even at that rate the voice track has stopped being the expensive line in an AI short.

What this means for people making work

In practice, for anyone making work this week, the bottleneck in AI voice has moved from the model to the page. When a synthetic voice can whisper, laugh and turn a line on a tag, the performance is only as good as the direction written into the script, and that is a craft skill rather than a subscription. Writers who can mark a scene the way a voice director would, with beats, pauses and intent, will get more out of v4 than people who paste in narration and hope. k_matsumaru is already thinking this way: he said the result made him want to turn his growing story world into a novel game rather than another video, which is a new format opening up rather than an old job disappearing.

The other opening is reach. At well under a dollar a pass, a festival short can now go out in several languages for the price of a lunch, which puts international distribution in reach of solo makers who could never book a dubbing studio. The same math applies to temp tracks, animatics and pitch reels, where a performed read used to mean hiring someone for a day.

Two cautions keep that honest. We have not listened to v4 ourselves, so nothing here is our own audio verdict; it rests on listener votes, one benchmark and creator posts. And some of the early buzz looks manufactured, though that is not proven. Several near-identical clone demos appeared within about 90 minutes of launch from accounts that advertise paid collaborations, and a staff member ran a credits giveaway for follows and reposts. The strongest signal is still the plain one: a director who already had characters and a story sat down in the middle of the night and got a scene he was happy to post unedited.

Timeline

• Date: Tue 3 Jun 2025. Event: Eleven v3 alpha launches with 70+ languages and audio tags; Source: ElevenLabs blog
• Date: Wed 20 Aug 2025. Event: Eleven v3 reaches the public API; Source: ElevenLabs blog
• Date: Wed 4 Feb 2026. Event: ElevenLabs closes a $500M Series D at an $11B valuation; Source: Round coverage
• Date: Early Jun 2026. Event: v4 previewed on stage at ElevenSummit Warsaw; Source: ElevenLabs event page
• Date: Wed 23 Sep 2026. Event: Alibaba releases Qwen-Audio-3.1, later first on the cloned-voice board; Source: Alibaba Cloud
• Date: Mon 28 Sep 2026, 14:01 UTC. Event: Eleven v4 and v4 Turbo launch with $22 and $11 promo pricing; Source: ElevenLabs blog and X
• Date: Mon 28 Sep, 17:48 UTC. Event: k_matsumaru posts a one-shot, unedited v4 voice drama; Source: X
• Date: Mon 28 Sep, 22:23 UTC. Event: Runway adds v4; Source: Runway on X
• Date: Tue 29 Sep 2026. Event: Cloning guide still recommends one to two minutes of audio; Source: ElevenLabs docs
• Date: Mon 12 Oct 2026. Event: Launch discount ends, v4 returns to $80 per million characters; Source: ElevenLabs pricing

What is still open

Nobody has published an independent test of a ten-second clone, and the reference clip length on the cloned-voice board is not stated. The company's pages disagree on whether instant clones need a spoken verification, and we did not make a clone to settle it. Two outlets report that clones made before v4 need retraining and that SSML break tags are disabled, but neither appears in the documentation, so we have left them out. How well tags hold across long narration in languages beyond English and Japanese is untested, and whether the template-style creator posts were paid is not proven.

Sources and reporting notes

Launch details come from the ElevenLabs blog post of Monday 28 September and its launch thread, read in full, and from the models documentation and Instant Voice Cloning guide, read on Tuesday 29 September. Prices come from the company's posts and API price page and the fal model page. Rankings come from the Artificial Analysis Provider Voice and Controlled Voice leaderboards, read on Tuesday. Latency testing is from Speko's post. Creator quotes are from k_matsumaru's posts on X, translated from Japanese, with impressions from the X API. Cost figures are our own calculation from list prices. We have not heard v4's output first hand.