ElevenLabs review: Eleven v4, pricing, cloning and privacy

This review is built from what could be checked on November 3, 2026: ElevenLabs' own site, pricing pages, documentation, help center, changelog, blog, status page, terms, privacy policy and Prohibited Use Policy; reporting by Tech.eu, Trending Topics, Engadget, MusicTech, Capitol News Illinois and Publishers Lunch; Artificial Analysis's public leaderboards and Vapi's Humanness Index; two comparisons by Tech Stackups; user reports on the App Store and Hacker News; and the alternatives' own pricing pages. Each piece of evidence says who produced it, and figures ElevenLabs gives about itself are labelled as its own.

What ElevenLabs is

ElevenLabs is an AI audio company founded in 2022 by Mati Staniszewski, its CEO, and Piotr Dabkowski (ElevenLabs). Users outside the EEA, the UK and Switzerland contract with Eleven Labs Inc. in New York (terms). It sells three things (about page, llms.txt):

• ElevenCreative, a web studio and mobile app for text to speech, voice cloning, voice design, dubbing, transcription, sound effects, music, and since September a video editor that also generates footage.
• ElevenAgents, a platform for voice and chat agents on the phone, the web and WhatsApp.
• ElevenAPI, the same models for developers, with Python and Node.js SDKs, an MCP server and a CLI.

The company says enterprise customers now bring 55% of its revenue and that its agents handle more than 15 million conversations a week (ElevenLabs). For most individual buyers, though, ElevenLabs is still the voice tool: paste a script, pick or clone a voice, download the audio.

What changed in 2026

• February 4: a $500 million Series D at an $11 billion valuation, led by Sequoia (ElevenLabs).
• May 7: API text-to-speech prices cut by up to 55%, transcription by up to 45% and agents by up to 20%, with prepaid pay-as-you-go top-ups replacing overage billing for new self-serve accounts (ElevenLabs).
• May: Illinois journalists, podcasters and voice actors sued ElevenLabs and eight other companies, including Amazon, Apple and Google, in Chicago under the state's biometric privacy law, alleging their voiceprints were taken from recordings to train voice models without consent (Capitol News Illinois, Publishers Lunch). The plaintiffs' law firm lists the cases as active (Loevy + Loevy).
• September 10: a multi-year licensing deal with Universal Music Group, ElevenLabs' first with a major label, starting with a fan remix platform that has no release date (ElevenLabs, Engadget, MusicTech).
• September 11 to 25: Music v2.5, Studio 4.0 and a free three-month plan for university students (ElevenLabs).
• September 28: Eleven v4 and v4 Turbo (ElevenLabs).
• September 30: a $300 million employee tender offer valued the company at $22 billion, double February's figure (Tech.eu). Trending Topics noted that a tender raises no new money for the company and comes with less disclosure than a funding round (Trending Topics).
• October 9 and 20: synthetic-voice detection for enterprise agents, then Eleven Realtime, two speech-to-speech agent models compatible with OpenAI's Realtime API (ElevenLabs).

ElevenLabs' own status page rates as major the disruptions to agent phone calls on October 6 and 8 and to dubbing on October 8; instant-clone creation also failed for a time on October 13.

Which model to use

From the API pricing page and the models documentation:

For transcription, Scribe v2 costs $0.22 an hour and the realtime version $0.39. The latency figures are ElevenLabs' own and differ between its pages: the launch post gives v4 Turbo about 150 ms to first speech, and the pricing page lists v3 Conversational at about 280 ms against 165 ms in the docs.

v4 changes how voices behave. Its documentation says it copies a source voice more faithfully, flaws included, so clean recordings matter more; that a clone speaking another language now gets a native accent instead of the speaker's, with no switch yet; that Style and Speed sliders and SSML are gone; and that voices made with Voice Design may sound less performative than on earlier models. ElevenLabs also warns that v4 is still being trained, so its output may shift and should be re-checked.

Plans, credits and what they buy

From the pricing page:

In the studio, speech costs one credit per character, and the plan table counts 10,000 credits as about 10 minutes, while transcription uses 330 credits a minute, music 900 and dubbing 2,000 to 10,000. The API works differently: usage is billed in dollars at the per-model rates above, so the same Creator plan buys about 275,000 characters of v4 through the API against 121,000 credits in the studio, and extra studio minutes cost $0.17 to $0.36 against about $0.08 through the API. Agents are billed by call minutes on a third tab (agents pricing).

Unused credits roll over for up to two months while you stay subscribed and are lost when you downgrade or cancel. For new self-serve accounts, extra usage now comes from prepaid top-ups of at least $5, which expire after 12 months, cannot be refunded and stop your service when they run out (pay-as-you-go docs). Refunds are possible within 14 days only if no credits were used (help center), and the terms say all sales are final.

Voice cloning, the Voice Library and payouts

An instant clone, from Starter, needs at least 10 seconds of audio (ElevenLabs recommends one to two minutes) and a confirmation that you have the right and consent to clone the voice (docs). A Professional clone, from Creator, is trained on 30 minutes to 3 hours of audio, takes 3 to 6 hours to fine-tune, and may only be of your own voice: you verify it by speaking, and someone else's consent is not enough (docs). A no-go list blocks clones that resemble prominent public figures (help center). Clones cannot be exported, so keep your original recordings (help center).

Only Professional clones can be shared in the Voice Library, where other users can pick them without spending a voice slot. Owners earn cash when paid users generate with their voice, set their own rate, a notice period of up to two years and optional Live Moderation, and are paid weekly through Stripe Connect once about $10 has built up (payouts docs). The payouts page says default rates depend on how long you commit to sharing, but the pages read do not publish them.

What public tests and users report

• Artificial Analysis's Provider Voice Arena, a third-party blind test. On November 3, 2026, Eleven v4 led 103 models with an Elo of 1334 and v4 Turbo followed at 1331, within each other's margin; then came Qwen-Audio-3.1-TTS-Plus (1301), Cartesia Sonic 3.6 (1293), Gemini 3.8 Flash TTS (1282) and Fish Audio Drama 3 (1272). On October 25, when GenGrades' list was written, Turbo was first. The arena's accent filters cover US and UK English only.
• ElevenLabs' own claim. Its launch post says about 75% of listeners preferred v4 in its blind tests against Cartesia Sonic 3.6, Inworld TTS-2 and two Gemini 3.8 voices (ElevenLabs).
• Cloned voices. On Vapi's provisional Humanness Index, run by a voice-agent platform, a cloned voice on Eleven v3 scored 98 against a real human's 100; v4 was not on it yet.
• Tech Stackups, two comparisons in October 2026. Against Gemini 3.8 Flash TTS, Eleven v4 won on numbers and names, a 2.5-minute narration and a dialogue, and lost on emotion and cloning. Against Voxtral TTS 3, it won every stock-voice test. In both, a clone of the reviewer's South African voice came out with a British accent.
• Transcription. Scribe v2 had a 2.2% word error rate on Artificial Analysis's index, fourth of the models listed.
• App Store users. Of the 100 most recent US reviews of the iPhone app, 33 gave one or two stars. The commonest complaints were credits spent on failed or partial generations and crashes after a mid-October update. The app still rates 4.8 from about 24,000 ratings (App Store).
• Hacker News, on Music v2.5's launch day. One user found the WAV export cleaner than Suno's, others hit outages and conflicting download limits (Hacker News).

Privacy, training and content rules

• Training. ElevenLabs uses data from non-enterprise accounts to improve its models until you switch off 'Improve the models for everyone'; the switch affects only new data (help center).
• Licence. The terms give ElevenLabs a perpetual, irrevocable, sub-licensable licence to your inputs and outputs, your voice included, to run and improve its services and build new products; it promises not to commercialise your voice on its own without permission. You keep the rights to your output.
• Storage. Personal data is stored in the United States, and data about your voice is kept no longer than three years after your last interaction (privacy policy). EU, India and Singapore residency is Enterprise-only (docs), as is Zero Retention Mode, which covers API calls but never the web app (docs).
• Use rules. The Prohibited Use Policy requires users to be 18, bans cloning a living person without explicit consent and recordings of minors, needs written approval for political campaigning, and forbids removing watermarks. Health data needs a signed BAA.
• Enforcement. ElevenLabs says it vets customers at sign-up, monitors with classifiers and human reviewers, bans serious violators and refers crimes to law enforcement (safety page). Generated audio carries a watermark that, by its own account, no other platform can yet detect (help center).

ElevenLabs lists SOC 2 Type II, ISO 27001, 27701 and 42001 and FedRAMP 20x Class A among its certifications (ElevenLabs). Under the terms for users outside the EEA, the UK and Switzerland, disputes go to binding arbitration with a class-action waiver unless you opt out within 30 days.

Who should not use ElevenLabs

• Anyone who wants a high-quality clone of someone else's voice: Professional clones are for your own voice only.
• Teams generating millions of characters a month on a tight budget; the alternatives below cost a fraction per character.
• Products whose voice must keep its accent across languages, which v4 changes by design.
• Self-serve users who need data kept outside the United States or not stored at all.
• Anyone who will not grant a vendor a perpetual licence to their recordings.
• Free users who want to publish commercially, and anyone under 18.

Alternatives and when they fit better

Prices from each product's own page, checked November 3, 2026; arena places from Artificial Analysis on the same day. GenGrades compares six options in its ElevenLabs alternatives list, written on October 25; since then v4 has overtaken v4 Turbo and Fish Audio's full Drama 3 release ranks sixth.

The Speechify API is also worth a look for its 500,000 free characters a month with commercial use. ElevenLabs' case against all of these is breadth and top-ranked quality: one account for voices, clones, transcription, dubbing, music and agents. If your bill is driven by volume, or you cannot accept its data terms, one of the alternatives fits better.