ElevenLabs v4: What’s New in AI Voice (2026)
ElevenLabs has shipped a new generation of its voice AI. Version 4 — actually two models, Eleven v4 and Eleven v4 Turbo — launched on September 28, 2026, and it arrived with an unusual amount of confidence behind it: the new flagship is currently sitting at #1 on the Artificial Analysis independent TTS Arena leaderboard, with an Elo score of 1,319.
But leaderboards don’t tell you what changed, what it costs, or whether it’s worth your attention. Let’s do the honest version.
Two models, two jobs
ElevenLabs split this release in two, and the split is deliberate.
Eleven v4 is the quality flagship. It’s the expressive one — richer emotional range, better prosody, the model you’d reach for when the voice is the product: narration, audiobooks, video voiceovers.
Eleven v4 Turbo is the latency play. It promises roughly 100ms median inference latency and about 150ms time-to-first-speech — numbers aimed squarely at voice agents and conversational AI, where a half-second pause makes the whole experience feel broken. Turbo is the fastest model ElevenLabs has shipped.
That separation matters. Older ElevenLabs models asked you to pick one model and live with its trade-offs. v4 lets you stop doing that: reach for quality when quality matters, reach for speed when the conversation does.
The feature list, translated
Official announcements pack a lot of jargon. Here’s what the additions actually mean in practice:
Inline expression tags. This is the headline feature for creators. You can now drop delivery instructions directly into your text — things like [laughs], or natural-language tags like “said angrily in a French accent” — and the model acts on them mid-read. Multiple tags can be stacked, so a single passage can shift tone several times. If you’ve ever generated a paragraph of TTS that sounded technically fine but emotionally dead, this is the fix ElevenLabs is betting on.
Better pronunciation. An improved pronunciation system with stronger IPA support — good news for anyone whose scripts include unusual names, brand terms, or mixed-language content. ElevenLabs also claims #1 on pronunciation robustness in independent testing.
Multi-speaker dialogue. Native support for generating conversations with distinct voices, which opens up dialogue-heavy content — skits, interviews, podcast-style scripts — without stitching separate generations together.
90+ languages. The release expands language coverage, with the biggest quality gains reported in Japanese, Brazilian Portuguese, Mandarin, and Cantonese. ElevenLabs calls this its most multilingual release yet.
10-second Instant Voice Clones. Cloning a voice now needs as little as 10 seconds of source audio. That’s convenience — and also the thing worth watching. The shorter the sample needed, the more the consent and disclosure conversation matters. ElevenLabs continues to require its usual voice-cloning verification steps. Use them, and don’t clone voices you don’t have the right to.
Your old voices still work. The existing library of 17,500+ voices carries over to v4, and legacy custom clones can be retrained onto the new models. If you’ve built a workflow around a specific voice, you don’t start over.
What changed vs v3
The honest summary: v3 was about making synthetic voices sound human. v4 is about making them directable. The expression tags are the clearest signal — ElevenLabs is shifting from “the voice sounds real” to “the voice takes direction,” which is the actual bottleneck for creators making long-form content.
The other big change is structural: splitting the lineup into a quality model and a latency model. That’s a mature-product move, and it tells you where ElevenLabs thinks the market is heading — high-fidelity narration on one side, real-time voice agents on the other.
What hasn’t changed: the voice still needs a good script. Tags can make delivery more expressive, but they can’t fix bad writing. Garbage in, expressive garbage out.
How it stacks up
The Artificial Analysis TTS Arena is an independent, community-voted leaderboard — blind human comparisons, thousands of them. Eleven v4 currently sits at Elo 1,319. For context, the next two are Cartesia’s Sonic 3.6 at 1,276 and Google’s Gemini 3.8 Flash TTS at 1,267.
A 40+ point Elo gap in this arena is a clear lead, not a rounding error. Leaderboards have limits — they measure what testers notice, and they can lag behind the newest releases from competitors — but as an independent check on ElevenLabs’ own claims, #1 with a margin is about as good as a launch-day signal gets.
Who it’s actually for
Voice agent builders — this is the core audience for v4 Turbo. If you’re building conversational AI, the latency numbers are the whole story: ~100ms inference, ~150ms to first speech. That’s near the threshold where conversation starts to feel natural.
Narrators and video creators — v4’s expression control is built for you. Audiobook producers, YouTubers, podcast makers: the ability to direct emotional delivery inline, in 90+ languages, with your existing voice library intact, is genuinely useful.
The merely curious — the free tier gives 10,000 credits per month (roughly 10 minutes of audio), and both new models are accessible on it. You can actually try v4 without paying anything.
The pricing math (honest version)
List pricing: $0.08 per 1,000 characters for v4 ($80 per million) and $0.04 per 1K for Turbo ($40 per million). Paid plans start at $6/month.
But here’s the angle that matters if you’re reading this in early October: there’s a 72% launch discount running until October 12, 2026, which drops pricing to $0.022 per 1K for v4 and $0.011 per 1K for Turbo. At those rates, Turbo costs roughly a penny per thousand characters — cheap enough to experiment with seriously.
A rough working estimate for budgeting: English narration runs around 800 characters per minute (very rough — it varies by voice and pacing). At launch-discount prices, that puts a minute of v4 Turbo audio at under a cent and v4 at under two cents. At list prices, roughly three and six cents per minute respectively. Neither is expensive by voiceover standards, but run the numbers against your actual volume before committing.
The discount window is short — October 12 is right around the corner. If you were already considering ElevenLabs, this is the week to test it. If you weren’t, a discount alone is never a reason to adopt a tool.
Where to use it
Both models are available across ElevenLabs’ main surfaces: ElevenCreative (the creative studio), ElevenAgents (the voice-agent platform), and the ElevenAPI for developers. API users call the new models the usual way — check ElevenLabs’ docs for the current model identifiers.
What to watch out for
A few honest caveats before you dive in:
Expression tags are a skill. Writing [laughs] in your script and getting a natural-sounding laugh are two different things. Expect a learning curve — prompting delivery is its own craft, and early attempts may sound worse than plain narration.
The free tier is a sampler. 10,000 credits a month is enough to evaluate, not to produce. A single long-form video can chew through that quickly. Go in knowing what the tier is for.
Discounts expire. The October 12 cutoff is real, and post-discount pricing is meaningfully higher. Do the math on what you’d actually spend at list price before building a workflow around discounted rates.
Voice cloning needs consent. Ten-second cloning is convenient and slightly alarming in equal measure. If you’re cloning client voices, talent voices, or anyone who isn’t you, get written permission. This isn’t just politeness — it’s increasingly a legal question, and ElevenLabs’ verification steps exist for a reason.
The bottom line
ElevenLabs v4 is a meaningful upgrade, not a rebrand: genuinely new directing capabilities, a proper latency tier, wider language coverage, and independent leaderboard results to back the quality claims. The 72%-off launch pricing through October 12 makes this the cheapest window to evaluate it — and the free tier means “evaluate” can mean actually using it, not just reading about it.
One caveat, since this matters: everything above is based on ElevenLabs’ official announcements and verified third-party reporting as of October 2026. I haven’t run hands-on listening tests on every feature, and TTS quality is subjective — the free tier exists precisely so you can judge the voices yourself. Use it.
Try ElevenLabs: if this article helped you, you can try the platform here — https://try.elevenlabs.io/haua0qrok9xl — (affiliate link: if you later buy a paid plan through it, I may earn a commission at no extra cost to you).