ElevenLabs

By Minel Gunesoglu, founder. I reconcile the review-platform numbers against the one public blind-listening leaderboard, and I read the terms of use myself so the licensing section is not guesswork. Prices collected 11 July 2026; the leaderboard row, the G2, Product Hunt and Trustpilot samples and the terms of use were all read on 12 August 2026. I re-check this page monthly; every dimension on this card carries a figure, which is unusual here, so expect the composite to move less than most. Last updated 18 August 2026.

ElevenLabs, the British text-to-speech and voice-cloning platform two researchers founded in 2022, carries one reputation almost everyone agrees on and one the measured evidence complicates. Take the reputation first: it sounds real. On Product Hunt a reviewer called the "voice quality of ElevenLabs is exceptionally professional and high-end" (Product Hunt, 2024), and that verdict recurs across G2 and Capterra as settled fact rather than a first impression. It is worth reading closely, though, because the praise almost always measures ElevenLabs against last-generation engines, Amazon Polly, Google Cloud TTS, older PlayHT, and not against the models winning blind listening tests today.

ElevenLabs voice quality: adored in reviews, eleventh of 93 on the blind Speech Arena

Put ElevenLabs on the one public blind-preference board I can verify and the ranking is more modest. On Artificial Analysis' Speech Arena, where listeners vote on same-text samples without knowing the vendor, ElevenLabs' best model, Eleven v3, sits eleventh of ninety-three by Elo (1,171), inside a tie band the board itself draws at tenth to eleventh; that is the Provider Voices view, read 12 August 2026. SpeechifyAI's Simba 3.2 leads at 1,231, with Google's Gemini 3.1 Flash TTS and Cartesia's Sonic 3.5 also rated above it. Eleventh of ninety-three is a strong result; it is not the outright best, which is precisely the claim the reviews keep making. That gap, adored in the review record and eleventh on the leaderboard, is the first of two splits running through everything ElevenLabs.

The negative reviews are not about competitive lag at all. They report concrete failures with the output itself rather than with what a rival can do, the one-star band of ElevenLabs' Trustpilot profile is where those sit, read 26 June 2026. No run from it is quoted here: the capture taken of that exact view does not contain the sentence this page previously attributed to it, and a quotation nobody can check is worth less than the count. Quality complaints and quality praise coexist in the same record, which is the honest starting point for a score.

ElevenLabs reviews split by who writes them: developers on G2, consumers on Trustpilot

The second split is louder. On G2, ElevenLabs holds 4.5 out of 5 across 1,172 reviews as read on 12 August 2026, with essentially nothing at one or two stars. On Trustpilot it holds 3.1 out of 5 across 1,093 reviews, read the same day, with 37% of them one-star against 48% at five. Same product, two nearly opposite verdicts, and averaging them into a single star rating would erase the reason they differ. The two platforms are not rating the same thing. G2's base skews developer and business buyers judging model quality and integration, the cohort most likely to renew; Trustpilot's skews consumers who arrived to complain about a charge or a cancellation. The score card below keeps them apart on purpose rather than blending them into one comfortable number.

Support divides the same way, and within days of itself. A five-star Trustpilot reviewer praised the "excellent tool, features and customer service" (Trustpilot, 7 July 2026); a one-star reviewer the same month found "no human agent to talk to, just an AI chatbot" (Trustpilot, 19 June 2026). Both are dated within weeks of each other, which reads as tiered or dispute-type support triage, not one company-wide standard.

Why ElevenLabs gets 1-star reviews: credits wiped the moment you cancel

The Trustpilot number has a specific cause, and it is not the voice. The one-star wave through June 2026 is overwhelmingly about a single billing policy: unused credits, including the rollover balance a subscriber already paid for, are wiped the moment a plan is canceled or downgraded. The complaint recurs across at least five separate reviews inside a ten-day window. "the second you cancel, every unused credit you've already paid for is wiped" (Trustpilot, 27 June 2026). "This company is a ripoff. They have now put 2 charges on my credit card." (Trustpilot, 24 June 2026). A third aimed at the paywall itself: "Slow and restrictive Pay-walled functions ( 90% ) for non-membership right upfront." (Trustpilot, 28 June 2026). This is the mechanism the cost section pulls apart, because the forfeiture rule and the per-attempt credit math are the two facts a buyer most needs before subscribing.

Is ElevenLabs safe? The robocall, the actors' lawsuit and the senator's inquiry

One more thread sits under the trust number, and it is not a review-platform gripe but a documented pattern. ElevenLabs was named as the likely engine behind the January 2024 fake-Biden robocall in New Hampshire, and it banned the account responsible. Two voice actors sued the company in 2024, alleging their voices, their "distinctive vocal timbres, accents, intonation, pacing, vocal mannerisms, and speaking styles" (Bloomberg Law, 29 August 2024), were cloned into default library voices without consent; that case became the first settlement among the roughly forty-eight AI-copyright suits filed to date, closing quietly in November 2025. In April 2026 a US senator opened an inquiry after the FBI reported $893M in AI-voice scam losses for 2025, and ElevenLabs pointed to "an extensive array of safeguards" (Forbes, 19 April 2026). The shape is real-world misuse followed by an after-the-fact policy response, three incidents across three years. On consent, the platform's own rule is that you may clone only your own voice, or someone else's from their verified account with explicit permission, and US state statutes (California's Civil Code 3344, New York's Civil Rights Law 50-51, Tennessee's ELVIS Act) require that consent independently. This is compliance context, not a workaround.

Is ElevenLabs worth it? The 85.0/100, and what crossing 80 does not mean

Start with the unusual thing about this card, which is not the total. Every one of the four dimensions carries a figure. On most cards here at least one row is blank, because a leaderboard has never listed the tool or a review platform will not serve us a rating, and a composite averaged over half a table moves a long way the moment the other half arrives. This one has nowhere left to surprise you, which is a better reason to lean on 80.6 than the number itself.

Two cells moved on 13 August 2026, in opposite directions, and the total barely registers what happened underneath it. The leaderboard reading was pulled forward from 11 July to 12 August because the voice category gained a second scored card whose capability comes off that same board on that same date, and two readings five weeks apart cannot honestly be set beside each other on a page that orders cards by their composite. The refreshed row went both ways at once: the Elo slipped from 1,174 to 1,171 while the placing improved from twelfth to eleventh and the field grew from eighty-nine entries to ninety-three, so capability eased 85.3 to 84.8. Usability travelled the other way, 90.2 to 91.8, on two repairs to the crowd figure. The Product Hunt component had been carrying 4.6 across 28 reviews; the copy of that page saved on 12 August reads 4.9 across 197, so a capture taken to give an old figure its bytes had spent six days quietly disproving the figure it was fetched to support. The Capterra component, roughly 4.75 across roughly 17 reviews, came out rather than in: that rating appears nowhere in the bytes this desk holds, because no copy of the page was ever kept, and being a small input does not excuse a number nobody can point at. Net movement, 80.4 to 80.6. Both of those figures have since been overtaken, and by a change to the ruler rather than to anything measured with it: on 15 and 17 August 2026 the scale itself was rebuilt, and the consumer record that had been sitting in a separate trust check was pooled into this row instead. Neither 90.2 nor 91.8 is the usability figure today. The paragraph is kept because it is what the card did on 13 August, and a page that quietly deletes its own history is harder to check, not easier.

The number has sat above 80 since 12 August, when the fourth dimension was rebuilt out of ElevenLabs' own terms of use and lifted the total across the line the site draws there. Read the one-word grade printed beside it for exactly what it is: a restatement of where 85.0 sits on a hundred-point scale, and nothing else. It is not a claim about how deep the evidence is, which is what the filled-cell count beside it reports, and it is not a verdict this desk formed separately from the arithmetic.

Read across and the marks explain themselves. Capability at 84.8 is the Arena Elo mapped onto our scale: high because the model genuinely is good, short of the top because eleventh is not first. Usability at 91.8 is where two of three consumer records now sit, and the third is the reason to read this row with the most caution on the card. G2's 4.5 across 1,172 and Product Hunt's 4.9 across 197 pool to a raw 92.3. Trustpilot's 3.1 across 1,093 no longer joins them, because its profile carries ElevenLabs' standing invitation to customers to review, the platform prints it, while nothing has been read about how the G2 and Product Hunt samples were gathered. Records are only averaged together where the way each was drawn is recorded as the same, and here it is not.

This is the softest number in the card, and it should be labelled as such. Every other movement on this site that day rested on a record whose provenance had been read on both sides or on a defect that could be pointed at. This one rests on an asymmetry: one record's gathering is known and the other two are simply unread. If the G2 and Product Hunt samples are opened and turn out to be invited as well, the three pool again exactly as they did before and this row returns to roughly 78.7, taking about three points of composite with it. Nothing about ElevenLabs would have changed in either direction. The 3.1 across 1,093 is not a discredited figure (it is on this page, dated, with the billing mechanism behind it pulled apart two sections up) and the gap between 4.5 and 3.1 on the same product is still the finding, not a rounding problem. Value at 93.5 reflects a real free tier plus that $6 Starter commercial floor. Commercial terms at 71.6 is the low row, and since 12 August 2026 it measures the contract rather than the customer record. Four positions in ElevenLabs' own document, of which exactly one is generous. You keep your Output; the free tier may not be sold; the indemnity runs against you; and the refund position adds nothing to what consumer law already gives you. The next section unpacks the indemnity, which is the position doing the most damage to that row and the one a professional most needs to read.

Two things sit outside all four rows. Uptime is one: how often a generation lands, and how fast it returns under load. That needs a controlled run I have not made, so the incident record appears below as context and is never converted into a figure. The consumer-trust record is the other, and being exact about the dates matters because this page has twice claimed more than it could. It left the commercial-terms row on 12 August 2026. It did not leave the card: from 15 August it was pooled into usability, where it was the single heaviest drag on the row, and this paragraph went on saying it fed no cell. It left for real on 18 August 2026, and the sentence is now true.

The limit this paragraph used to describe is also gone. A rating under half marks on a sample this deep once put a hard lid over the whole composite, whatever the four rows underneath it said. Nothing works that way now: evidence either lands in a cell or it does not reach the number at all. This record lands in no cell and runs through most of this page. How each dimension is built is public.

Does ElevenLabs own my cloned voice? The perpetual license, and who indemnifies whom

This is the section every competing review skips, and it is the one a professional should read before uploading a voice or shipping client work. ElevenLabs' terms are plain on the first question: you keep the rights to your Output. What the marketing does not foreground is the license attached to it. By generating on the platform you grant ElevenLabs a perpetual, irrevocable, royalty-free, worldwide, sub-licensable license to use that content to run the service, and if you train a cloned voice, the same broad license covers the uploaded voice audio (an opt-out for training exists in account settings). ElevenLabs does commit that it "will not commercialize your voice on a standalone basis without your permission," which is a real guardrail, but it is narrower than owning the model of your own voice outright.

The sharper surprise is indemnification, and it runs opposite to what most buyers assume. In these consumer terms the user indemnifies ElevenLabs against third-party IP claims arising from what the user generates; ElevenLabs does not indemnify the user in return, and the obligation reads identically for Free and paid accounts. For a solo creator that is usually an abstract risk. For a design team or agency putting synthetic voice into client deliverables it is a line to price in before signing, and it carries the heaviest weight of the four positions behind the commercial-terms row above, which is why that row is the lowest on the card.

On the cloning itself, ElevenLabs gates the good version behind a plan. Instant Voice Cloning, built from a couple of minutes of audio, ships on Starter; the more stable Professional Voice Clone used for narration-grade consistency needs Creator ($22/mo) or above and a longer training sample. A "voice cloning: yes" checkbox on a comparison grid hides that split entirely.

ElevenLabs pricing and credits explained: why free-tier audio cannot be sold

ElevenLabs is billed in credits, not minutes, so the figure that governs a purchase is how far a monthly credit allowance stretches. Pricing verified against elevenlabs.io/pricing on 11 July 2026:

Two rules turn that table from marketing into a budget. First, the Free plan carries no commercial rights: its 10,000 credits (roughly ten minutes of text to speech) are fine for testing, but ElevenLabs' terms restrict Free output to non-commercial use, so a client voiceover generated there cannot legally be monetized. Commercial use, and with it the right to sell what you make, begins at Starter. I leave out the attribution required on Free claim circulating in other reviews: ElevenLabs lists "no attribution" as a paid-tier feature, which implies but does not state a free-tier attribution rule, so I will not assert one. One point the is it free searches keep conflating: ElevenLabs runs no time-limited trial. The Free plan is a standing tier you can sit on indefinitely at 10,000 credits a month; it just cannot be used to make money.

Second, credits are charged per generation attempt, not per usable clip. A regeneration to fix one mispronounced word costs the same as the first take, which is the documented root of the credits vanish faster than the minutes I made complaint: my shorthand for the pattern, not anybody's words. One Product Hunt reviewer put the general version plainly: the "system actually uses credits, not unlimited monthly usage as advertised" (Product Hunt, c. October 2025). Different jobs draw at very different rates: text to speech runs 1 credit per character (0.5 on the Flash and Turbo models), while speech to text costs 330 credits a minute, the sound-effects generator 200 credits a generation, the voice changer 1,000 credits a minute, and automatic dubbing 13,500 credits a minute on paid plans (22,000 on Free).

Then the cancellation trap, which is the credit-forfeiture wave from the review record restated as a rule. Subscription credits, including any rolled-over balance (unused monthly credits roll over up to two months, capped at three times the monthly quota), are forfeited immediately when you cancel or downgrade. Separately bought pay-as-you-go credits behave the opposite way: cancellation does not touch them, but they expire twelve months after purchase. Buyers routinely conflate the two, cancel expecting to keep a paid-for balance, and lose it. Spend a subscription balance down before you change plans.

ElevenLabs API pricing per character, and the uptime nobody scored

ElevenLabs' API is the part developers like most, and the reviews back that: one called it "clean and well documented, which made integrating it into our workflow pretty painless" (Product Hunt, c. May 2026). API pricing tracks the same per-character logic as the dashboard, roughly $0.05 per 1,000 characters on Flash and Turbo against about $0.10 on the Multilingual models, and one onboarding surprise is worth flagging: API calls draw from the same monthly credit pool as the web app, not a separate developer allowance.

The caveat is one the score card cannot reach at all. Developer reviews measure how pleasant the API is to integrate, not how often it is up. The independent monitor StatusGator has been collecting ElevenLabs outage data since May 2025, and its own summary line, captured on 12 August 2026, reads "more than 224 outages that affected ElevenLabs users" over that period, including five in June and July 2026 spanning dubbing failures and elevated error rates. An earlier version of this sentence said 212, and that figure is withdrawn (not because it was wrong, but because no capture of the page holding it was ever kept, so it cannot be reopened; the string 212 appears nowhere in the bytes we hold. A running counter with no dated capture behind it is a number that only its author can check. Individual entries rotate off the page as the window rolls, and two the 12 August capture still carries are "Speech Engine Conversation Failures" (14 July 2026) and "Elevated error rates on TTS, STT APIs" (3 August 2026). A correction belongs with that list: an earlier version of this passage put an incident titled Website Slowness and API Request Latency in quotation marks against this source, and that string is in none of the bytes on file. StatusGator server-renders no incident history beyond the current window and links no archive, so a title that has already rotated out can be shown neither way) and a quotation that cannot be shown does not stay a quotation. The recurring is ElevenLabs down searches line up with that log. It is real context, but it is not a controlled uptime test I ran, so it sharpens this warning without becoming a scored number.

ElevenLabs alternatives, and the one thing each of them does better

ElevenLabs is scored on the same instruments as every tool in our voice-generator rankings, and each name below beats it on one specific, dated axis rather than in general. No head-to-head page follows for any of these pairings: that format was retired here on 11 August 2026, and one product is never set against another on this site. An earlier version of this sentence promised one; it is withdrawn. What each of these tools gets instead is its own card on the same four dimensions, which a reader can hold against this one. The table below is the whole of the cross-tool claim: a dated fact per name, and nothing about which product is better in general.

Set the two splits side by side and ElevenLabs resolves, as of 12 August 2026, into a recommendation with conditions attached. The voice is genuinely excellent and the tooling around it (voice cloning, dubbing, a sound-effects generator, a clean API) is deep, which is why developers and creators rate it 4.5 on G2 and why capability and value do most of the work holding the 85.0 up. But the one blind leaderboard puts its best model eleventh of ninety-three, not first; the company's own consumer-trust record sits at 3.1 on the strength of credit forfeiture on cancel; and the terms hand ElevenLabs a perpetual license while handing you the indemnification bill, which is the contract row's 71.6 in one sentence. So the answer to "is ElevenLabs worth it" depends on which buyer you are. If you want the best-known voice and will manage the per-attempt credit math and the cancellation rule deliberately, start on the free tier, then buy Starter for commercial rights. If you are putting synthetic voice into client work, read the license and the indemnification clause before you commit. The reputation is earned; it is just not the whole measured picture, and this page is re-checked monthly as both the leaderboard and the terms keep moving.

Minel Gunesoglu, founder of Vouch, built this score and wrote the page. Connect on LinkedIn, or read what goes into the audit. Every rating, rank, and price here is dated at the point it appears and re-checked monthly. Disclosure: some outbound links on this site may be affiliate links; a paid submission only ever buys a faster review, never a listing or a better score, both fixed by the methodology alone.