Eleven v4 and v4 Turbo: What ElevenLabs Just Shipped

Disclosure: the ElevenLabs links on this page are affiliate links — ElevenLabs pays us a commission if you sign up through them, at no cost to you. It is the only company on this page that does. Everything below is drawn from ElevenLabs' announcement and independent launch coverage, named where it is used.

What launched

ElevenLabs announced both models together on 28 September 2026, in a post titled "Eleven v4: Our most expressive text-to-speech AI model yet" and a thread on its X account. The launch was covered the same day by TechCrunch, Unite.AI and VKTR, among others. The headline facts are consistent across all of them:

• Two models, one family. Eleven v4 for expressive output; Eleven v4 Turbo for real-time use.
• More than 90 languages, per TechCrunch's report and ElevenLabs' own announcement.
• Voice cloning from as little as 10 seconds of audio, ElevenLabs said, with speaker identity held more reliably across long-form content.
• Available immediately in ElevenAgents (the agent platform), ElevenCreative, and through the API — including on the free tier, according to Unite.AI.

What is actually new: direction, not just a better voice

ElevenLabs describes v4 as built on a new architecture that reads tone, pacing, emotion and context from the text itself. In practice, the more useful change for builders is how you steer it. According to launch coverage, you can direct delivery in plain language and drop inline audio tags into the script — the examples ElevenLabs gave include a laugh, a line said angrily in a French accent, light rain in the background and a phone buzzing.

If you have used the audio-tag controls on Eleven v3, this is the same idea taken further. The difference that matters is where it is available: expressive control used to live on a model ElevenLabs itself said was unsuited to real-time use. With v4 Turbo, a version of that control reaches the live-agent tier.

The Turbo numbers, read carefully

Launch coverage quotes ElevenLabs at roughly 100ms median inference latency for v4 Turbo and roughly 150ms median time to first speech. Those are two different measurements, and it is worth keeping them apart: inference latency is how long the model takes to produce audio; time to first speech is closer to what a caller actually experiences.

For comparison, ElevenLabs' model documentation lists Flash v2.5 — until now the default for live agents — at around 75ms model inference time. The figures come from different publications, so treat the comparison as indicative rather than exact. But the honest reading is that v4 Turbo is not a speed upgrade over Flash v2.5. It is an expressiveness upgrade at speeds that still work for conversation.

That is a reasonable trade for a lot of agents. A support line or an AI tutor that sounds flat loses people faster than one that takes an extra few tens of milliseconds to reply. It is a bad trade for anything where every millisecond is already spoken for.

Which model should you use?

Who v4 is wrong for

This is a good release, but it is not an automatic upgrade. Hold off if any of these describe you:

• Your agent is latency-bound and already working. On the published figures Flash v2.5 is still the faster model. If users are not complaining about how your agent sounds, there is nothing to fix.
• You need identical output every run. More expressive models vary more between generations. For compliance scripts, legal disclaimers or anything that gets audited word-for-word, predictability is worth more than emotion.
• You run a production pipeline with pinned voices. A new model can change how a cloned or designed voice sounds. Re-generate a sample set and listen before you swap anything in production.
• You are cost-sensitive at high volume. Paid pricing for v4 was reported inconsistently at launch. Confirm your per-character cost on ElevenLabs' pricing before you move a large workload.

How to try it today

Because v4 is on the free tier, the cheapest test is a real one: take three lines from your own product — a greeting, an apology and something with a number in it — and generate each on Flash v2.5, v4 Turbo and v4. Numbers and apologies are where text-to-speech tells on itself, and ten seconds of listening will tell you more than any benchmark.

If you call the API, switch models with the model_id parameter. We are deliberately not printing the v4 model ID here: copy it from ElevenLabs' models documentation rather than from a blog post, so you get the current value instead of one that may have changed since launch.

If you are building your first voice agent rather than upgrading one, start with the free ElevenLabs course below — lesson one walks through a working agent end to end, and v4 Turbo drops straight into the same setup.

Where this leaves the rest of our ElevenLabs coverage

Our text-to-speech roundup has been updated for v4. The full ElevenLabs review and the voice agent guide still describe the v3 and Flash v2.5 lineup, which remains available — v4 adds to the range rather than replacing it, as far as anything published at launch says. Want the competitive picture? See ElevenLabs vs PlayHT.

Ready to build? Try ElevenLabs free — v4 and v4 Turbo are both on the free plan.