ElevenLabs v4 vs. v4 Turbo: What’s New and Which Model to Choose

ElevenLabs’ September 28, 2026 launch splits its latest speech generation into two options: Eleven v4 for expressive, produced audio and Eleven v4 Turbo for real-time applications. Both support more than 90 languages and voice cloning, but their intended uses differ—and launch claims about quality and speed are not a substitute for testing with your own content. 51117

Eleven v4 vs. Eleven v4 Turbo at a glance

	Eleven v4	Eleven v4 Turbo
Designed for	Content creation, audiobooks and character voiceovers where quality is the priority 17	Real-time use, including conversational agents and interactive voice experiences 17
Performance emphasis	Expressive delivery and control over a script 511	Lower-latency speech generation 111
Reported latency	No comparable figure cited in the available sources	About 100 ms median inference latency; about 150 ms to first speech, according to ElevenLabs 11
Language support	More than 90 languages 111	More than 90 languages 1
Availability	ElevenAPI, ElevenAgents and ElevenCreative 210	ElevenAPI, ElevenAgents and ElevenCreative 210

Choose v4 when delivery, emotion and script control matter most. Choose Turbo when response speed is central to the application. The available documentation describes Turbo as retaining high quality while reducing latency, but does not provide a like-for-like quality comparison across every use case. 17

What’s new in v4: expression, cloning and longer scripts

ElevenLabs describes v4 as using a new architecture intended to give more control over tone, pacing, emotion and character. The public launch materials describe those goals but do not provide enough technical detail to independently assess the architecture itself. 510

The model also expands the use of inline audio tags, which let creators guide how a line is delivered. ElevenLabs introduced inline tags with v3; TechCrunch reported that v4 supports stacking tags and following their sequence. 5 That offers more direction inside a script, but the result still needs to be checked against the intended performance.

Both new models support more than 90 languages. ElevenLabs says an Instant Voice Clone can be created from as little as 10 seconds of audio. Professional Voice Clones, which the company says were not supported in v3, return in v4. 1511

For longer scripts, ElevenLabs says its context stitching feature helps keep pacing and delivery steady across a script of any length. That may be useful for audiobooks and other long-form work, but it is a product claim—not independent evidence that every voice remains consistent in every project. 11

Turbo’s latency figures—and what they measure

ElevenLabs reports roughly 100 ms median inference latency for v4 Turbo and roughly 150 ms median time to first speech. These are different measurements: one describes inference latency, while the other refers to the wait before speech begins. Neither figure, on its own, establishes how quickly a complete voice-agent interaction will feel to a caller. 11

A live interaction also depends on the rest of the system, including processing the caller’s audio, generating the agent’s response and delivering it over the network. So treat the published latency figures as useful model-level specifications, then measure response times in the application you plan to deploy.

How strong is the evidence on quality and competition?

A launch report says Artificial Analysis’ Provider Voice Arena leaderboard placed Eleven v4 at No. 1. That is a result for v4 on a particular evaluation; it does not establish that v4 leads on every language, voice, long-form task or live-agent setup. 6 The cited ranking should not be applied to Turbo: a separate sourced public benchmark for v4 Turbo was not identified in the available benchmark information. 18

Cartesia and Inworld are among the other providers discussed in coverage of real-time voice tools. Their existence makes direct testing important: compare the voices and latency using the same scripts, network conditions and agent workflow rather than relying on vendor latency claims alone. 19

What the launch says about ElevenLabs’ position

The model split reflects a broader product strategy: serve creators who want controlled, expressive narration while also targeting companies building real-time voice agents. That is a product-positioning signal, not proof that either model has won its market.

The company has reported more than $500 million in annual recurring revenue in early 2026, and its February 2026 funding round valued it at $11 billion. 3233 TechCrunch later reported that the company was pacing at $600 million in annual recurring revenue and that it was reportedly valued at $22 billion; the reported valuation should not be confused with a confirmed new funding round. The same report discussed IPO timing as an ambition, not an announced listing. 28 TechCrunch also described Decagon, a former ElevenLabs customer, as a competitor—an example of competition extending into customer-facing voice products. 28

For teams evaluating the models, the practical takeaway is straightforward: start with the use case, then test the specific voice and workflow. Compare v4 and Turbo on the same script or agent task, measure end-to-end latency, and verify that cloning and long-form consistency meet your needs. The launch establishes a clear quality-versus-speed choice; it does not settle which model will perform best for your project.