ElevenLabs
ElevenLabs has the most expressive AI voice model in the market (Eleven v3) and the broadest language coverage among major TTS providers, but the Free plan is a non-commercial sandbox requiring attribution, Professional Voice Cloning is limited to your own voice even with consent, concurrency is a plan-level ceiling that credits cannot raise, v3 is explicitly not for real-time use, and Cartesia matches its naturalness at roughly half the price and a fraction of the latency. It is a quality-first purchase with rights, throughput, and legal constraints that determine whether it is usable for a given workflow.
What is ElevenLabs?
ElevenLabs is an AI voice platform, not a single product. Its official structure splits into three families that share the same six headline prices but meter different things: ElevenCreative for voiceover, dubbing, music, and content production; ElevenAgents for conversational voice agents billed per call minute; and ElevenAPI for developers embedding speech into products, billed per character with plan-based concurrency. A team that uses both marketing voiceover and a customer-support voice agent is metered twice—once in creative credits and once in agent minutes—before the LLM and telephony costs that sit outside ElevenLabs' pricing page. Knowing which family a purchase belongs to is the first budgeting decision, and it determines which rate card applies.
The product's reputation rests on voice quality. Eleven v3, generally available since February 2, 2026, is the expressive flagship: it supports inline audio tags for emotion and delivery, speaks 70+ languages, and at GA reduced structured-text errors from 15.3% to 4.9% across a 27-category, 8-language benchmark, with chemical formulas dropping from 45.6% to 0.6% and phone numbers from 16.9% to 0.6%. On independent naturalness benchmarks, eleven_v3_conversational scores an Elo of 1590, among the highest in the market. That quality is the reason many teams choose ElevenLabs and the reason competitors benchmark against it.
The pricing architecture is where the product becomes more complicated than its marketing. Credits are shared across all products in a single monthly pool. One credit is roughly one character of standard Multilingual v2 speech, so a Creator plan's 121,000 credits buy about two hours of TTS—or about two hours of voice-changer audio, but not both. Dubbing is the fastest credit consumer at 2,000–10,000 credits per minute; a Creator plan covers roughly 12 minutes of watermark-free Dubbing Studio video. Unused credits on paid plans roll over for up to two months, capped at three times the monthly quota, but downgrading or cancelling forfeits the balance. The May 2026 price cut of up to 55% reset the rate card, and many third-party guides still show the old figures—budgeting from a stale source is a real risk.
Three structural constraints define the boundary of the product. The first is rights: the Free plan has no commercial license, and every published non-commercial output must credit ElevenLabs. Commercial use begins at Starter. The second is concurrency: the plan, not the credit balance, determines how many simultaneous generations are allowed (2 on Free, 15 on Scale and Business), and top-ups do not raise it. The third is cloning: Professional Voice Cloning requires identity verification and is limited to the user's own voice, even when another person has consented. These are not edge cases; they determine whether ElevenLabs can be used for a given project at all.
How does ElevenLabs work?
A user selects a model and either types text, uploads a script, or calls the API. The text is converted to speech, with model choice determining the trade-off between expressiveness, latency, and cost. Eleven v3 ($0.10 per 1,000 characters on the API, 5,000-character request cap) is the expressive choice for narration and dialogue; Flash v2.5 ($0.05 per 1,000 characters, ~75ms latency, 40,000-character cap) is the real-time choice; Multilingual v2 sits between them for long-form stability across 29 languages. Audio tags placed inline in the text—[whispers], [laughs], [sighs], [british accent]—are read as performance direction rather than spoken words, and the model applies them to the surrounding delivery.
On the creation side, ElevenCreative provides a studio for voiceover, dubbing, music, and sound effects, with the credit pool determining monthly capacity. On the agent side, ElevenAgents bills per call minute (15 included on Free, 12,375 on Business at $0.08/minute) with separate concurrency for calls, and the underlying LLM and telephony carrier are billed independently—ElevenLabs adds no telephony fee of its own, but the carrier's per-minute rate and the model's per-token cost are real lines in the budget. On the developer side, the API exposes speech, streaming speech, and separate dialogue endpoints for multi-speaker audio, with the same plan tiers governing concurrency and feature access.
The output is audio, but the deliverable depends on the family: a rendered voiceover file, a live voice agent call, or streaming speech inside a product. Rights travel with the plan, not the file: content generated on a paid plan can be used perpetually for commercial purposes even after the subscription ends, but content generated outside a paid subscription—before starting or after ending—cannot be used commercially. This makes the billing cycle boundary a rights boundary as well, which is easy to overlook when pausing a subscription between projects.
Use Cases
- Expressive narration, audiobooks, and character dialogue where emotional range and audio tags matter more than real-time latency
- Multilingual voiceover and dubbing across 70+ languages, accepting that dubbing consumes credits fastest and quality outside English is lower
- Real-time voice agents and conversational IVR using Flash v2.5 or v3 Conversational, with LLM and telephony costs budgeted separately
- Instant Voice Cloning of one's own voice for consistent branded content, or Professional Voice Cloning of one's own voice for higher fidelity
- API integration of TTS into products, with plan tier determining output quality (44.1kHz PCM from Pro) and concurrency
- Sound effects and music generation within the same credit pool as speech, with the caveat that music and dubbing drain the pool quickly
Pricing
Pricing model: Freemium
| Plan | Price | Billing | Included usage | Notes |
|---|---|---|---|---|
| Free | $0 | per month | 10,000 credits per month. Roughly 10 minutes of Multilingual v2/v3 speech or 20 minutes of Flash/Turbo. 2 concurrent Multilingual requests, 4 concurrent Flash requests. No commercial license. Any published non-commercial output must credit ElevenLabs with a link to elevenlabs.io. No Professional Voice Cloning. No API access to 44.1kHz PCM. | — |
| Starter | $6 | per month | 30,000 credits per month. Commercial license. Instant Voice Cloning. 3 concurrent Multilingual requests, 6 concurrent Flash requests. API access with standard output formats. This is the practical minimum for any monetized use. | — |
| Creator | $22 | per month | 121,000 credits per month. Professional Voice Cloning. 5 concurrent Multilingual requests, 10 concurrent Flash requests. The advertised $11 is a first-month promotion, not the recurring price. | — |
| Pro | $99 | per month | 600,000 credits per month. 44.1kHz PCM output via API. 10 concurrent Multilingual requests, 20 concurrent Flash requests. | — |
| Scale | $299 | per month | 1,800,000 credits per month. 3 workspace seats. 3 Professional Voice Clones. 15 concurrent Multilingual requests, 30 concurrent Flash requests. | — |
| Business | $990 | per month | 6,000,000 credits per month. 10 seats. 10 Professional Voice Clones. Low-latency TTS marketed from 5¢ per minute. 15 concurrent Multilingual requests, 30 concurrent Flash requests. | — |
| Enterprise | Custom | annual | Custom credits and concurrency. DPAs, SLAs, BAAs for HIPAA, SSO, Zero Retention Mode. Training on customer content is off by default on Enterprise; custom terms available. | — |
Common Limitations
- The Free plan cannot be used commercially — this is the single most important fact about the product · ElevenLabs' Free plan carries no commercial license. Any content generated on Free and published—whether a YouTube video, a client deliverable, a game asset, or a podcast—must credit ElevenLabs with a link to elevenlabs.io, and cannot be monetized. The real entry price for any commercial use is Starter at $6/month, not $0. This is a harder boundary than many freemium tools: it is not a volume limit that a top-up can solve, but a rights limit that requires a subscription change. Teams that prototype on Free and then ship without upgrading are in breach of the terms, and the attribution requirement applies to all non-commercial public output as well. ElevenLabs' own help documentation states the free plan 'does not include a commercial license and may not be used for any commercial purpose,' and reviews consistently flag this as the gating fact that determines whether the product is usable at all for professional work.
- Concurrency is a plan-level ceiling that credits cannot raise · ElevenLabs meters concurrency per plan, not per account balance: 2 simultaneous Multilingual requests on Free, 3 on Starter, 5 on Creator, 10 on Pro, 15 on Scale and Business. Exceeding the limit does not simply bill overage—requests queue, and on the Free tier they are throttled hard. For a batch dubbing job or a product that serves multiple listeners simultaneously, concurrency, not credits, is the binding constraint. A team can have thousands of credits remaining and still be unable to serve more than two simultaneous speech generations. Pay-as-you-go top-ups raise the character budget but never the concurrency limit, so a burst of concurrent users cannot be bought out of on a low tier. The practical consequence is that live applications and batch pipelines often need to upgrade for throughput before they exhaust their credit allowance, which is the opposite of the intuition that plan choice is about volume.
- Professional Voice Cloning is restricted to your own voice — even with explicit consent, you cannot clone someone else · ElevenLabs distinguishes Instant Voice Cloning (available from Starter) from Professional Voice Cloning (available from Creator). The Professional tier requires identity verification, and ElevenLabs' documentation and third-party legal analysis both state that you can only create a Professional clone of your own voice. A request to clone someone else is prohibited even if that person has given written consent. The Voice Library appendix adds that user voice models must be based on the user's own voice and that modified versions of one's own voice are also restricted. This is stricter than the colloquial understanding of 'consent-based cloning' and rules out the agency use case of cloning a client's voice actor through ElevenLabs' own Professional pipeline. The only compliant path for third-party voices is the licensed talent route or the Iconic Marketplace for celebrities, not self-serve cloning. At least 12 US states have voice-cloning consent laws, and ElevenLabs has suspended accounts for violations, so the restriction is enforced rather than advisory.
- Default training-on-data for non-Enterprise tiers, with a three-year retention floor · ElevenLabs' terms grant the company a perpetual, irrevocable, royalty-free, worldwide license to use voice data and content to train models and create derivative works. Free and Growth-tier users have training data use on by default; opting out requires a manual toggle in the Data use menu and applies only prospectively. Raw voice recordings are deleted after three years of inactivity, but models and derivatives created from them are retained permanently. Enterprise customers get the inverse default: training is off unless they explicitly opt in, with DPAs, BAAs, and Zero Retention Mode available. The practical consequence is that a Free or self-serve paid user who does not find and change the setting is on a path where their voice data may be used for model improvement indefinitely, and the three-year deletion clock does not remove derivatives already created. For teams handling sensitive audio, the configuration change is mandatory, and contractual protection requires Enterprise.
- Eleven v3 is explicitly unsuitable for real-time use and caps requests at 5,000 characters · Eleven v3, the expressive flagship model, carries two documented constraints that its marketing emphasizes less than its emotional range. First, the maximum characters per request is 5,000—roughly five minutes of speech—so long-form pipelines that assume a larger per-request limit will fail at the API boundary. Second, ElevenLabs' own materials describe v3 as 'not suitable for real-time use.' The model is positioned for offline, expressive narration and dialogue, not for live voice agents. Real-time applications should use Flash v2.5 (~75ms latency, 40,000-character limit) or v3 Conversational (~280ms), which trade some emotional range for responsiveness. Independent field reports also note that v3's audio tags can occasionally flip a speaker's accent entirely, that tag effectiveness varies by voice and context, and that one reviewer described the overall experience as still feeling like a beta—with an ElevenLabs staff account acknowledging the issues were being worked on. v3 is the best-sounding ElevenLabs model, but it is not a drop-in real-time engine and not a finished control surface.
- Cartesia is faster and cheaper on the same unit, and latency leadership has moved away from ElevenLabs · On the independent Speko benchmark, ElevenLabs eleven_v3_conversational scored a naturalness Elo of 1590 with first-audio p50 latency of 266ms; Cartesia sonic-3.5 scored 1574 at 121ms, at the same $50 per 1M characters. Eleven v4 Turbo is vendor-stated at roughly 150ms median, still behind Cartesia's sub-90ms figures. Independent analysis places Cartesia Sonic 3.6 at roughly $49 per 1M characters against $100 per 1M for Eleven v3 as tracked by Artificial Analysis, a roughly 2:1 price gap. ElevenLabs' advantage is concentrated in expressive, offline narration quality and language breadth (70+ languages on v3), where it remains a default choice. For live voice agents where time-to-first-audio and per-character cost dominate, the competitive set has moved to Cartesia, Deepgram, and Flash v2.5, and ElevenLabs' own v3 is not the model to use. The honest positioning is that ElevenLabs is a quality leader in one lane and a premium-priced follower in another—not the fastest or cheapest option across the board.
- Ongoing legal and political scrutiny over misuse and training data · ElevenLabs faces two distinct legal pressures that buyers should weigh. In May 2026, a set of class actions filed in the Northern District of Illinois—including Amer v. Eleven Labs Inc., 1:26-cv-05437—alleged that ElevenLabs ingested broadcast journalists' voice recordings into its training pipeline and extracted BIPA-protected voiceprints without consent, seeking $1,000 per negligent violation or $5,000 per intentional violation. ElevenLabs filed a motion to dismiss in August 2026, arguing the plaintiffs did not plausibly allege a BIPA-protected voiceprint. Separately, in April 2026, Senator Maggie Hassan sent a letter to ElevenLabs, LOVO, Speechify, and VEED pressing them on scam prevention, following an FBI report that AI-related scams caused $893 million in reported losses in 2025. ElevenLabs responded that it maintains an extensive array of safeguards. The company also reached an agreement with Brazil's Superior Electoral Court in August 2026 to prevent voice cloning during elections. These are not evidence that the product is unlawful, but they are active, material risks: the training-data litigation directly challenges the ingestion model, and the scam scrutiny targets the misuse vector that any voice-cloning platform enables. Enterprise buyers should ask for indemnification and audit terms rather than assume the standard self-serve posture covers these exposures.
Key Features
- Eleven v3 expressive TTS with inline audio tags and 70+ languages
- Flash v2.5 real-time TTS at ~75ms latency with 40,000-character requests
- Instant Voice Cloning from Starter; Professional Voice Cloning from Creator (own voice only)
- Speech-to-text, sound effects, music, and dubbing within the same credit pool
- ElevenAgents conversational voice agents billed per call minute with separate concurrency
ElevenLabs vs. Cartesia and OpenAI TTS
This comparison uses official pricing pages, model documentation, and independently published benchmarks available as of November 2026. It focuses on the structural differences that determine where each provider is the right fit: expressive quality versus latency, per-character cost, concurrency behavior, and licensing boundaries. ElevenLabs is the quality-first choice for offline, expressive narration and broad language coverage; Cartesia is the latency-first choice for real-time agents at a lower per-character price; OpenAI TTS is the ecosystem-integrated choice for teams already standardized on OpenAI models, with fewer controls and a simpler contract surface.
| Tool | Positioning | Key differences | Best fit | Pricing note |
|---|---|---|---|---|
| ElevenLabs Official pricing and plan comparison Official help: commercial use and attribution Independent benchmark (Speko, naturalness and latency) | The quality-first AI voice platform, with Eleven v3 as the expressive flagship (70+ languages, inline audio tags) and Flash v2.5 (~75ms) for real-time use, sold across three product families on six shared dollar tiers. | Highest naturalness Elo among the three on independent benchmarks (1590 for v3 Conversational), and the broadest language coverage. But the Free plan is non-commercial and requires attribution; commercial use starts at $6/month. Professional Voice Cloning is own-voice-only even with third-party consent. Concurrency is plan-gated (2 on Free, 15 on Scale/Business) and cannot be raised by top-ups. Eleven v3 is explicitly not for real-time use and caps requests at 5,000 characters. Training on voice data is on by default outside Enterprise, with a three-year retention floor. Cartesia matches naturalness within ~16 Elo points at roughly half the per-character price and about half the latency. | Expressive offline narration, audiobooks, character dialogue, and multilingual voiceover where naturalness and language breadth matter more than real-time latency or per-character cost—provided the buyer is on a paid plan and has cleared the cloning rights boundary. | Free $0: 10k credits, no commercial license, attribution required, 2 concurrent. Starter $6: 30k credits, commercial license, Instant cloning, 3 concurrent. Creator $22: 121k credits, Professional cloning, 5 concurrent. Pro $99: 600k credits, 44.1kHz PCM, 10 concurrent. Scale $299: 1.8M credits, 15 concurrent. Business $990: 6M credits, 15 concurrent. API $0.04–$0.10/1k chars. ElevenAgents $0.08/min plus LLM and telephony. |
| Cartesia Independent benchmark (Speko, Sonic 3.5 latency and naturalness) Independent analysis (2:1 per-character price gap) | A latency-first TTS provider built on State Space Model architecture, with Sonic models designed for real-time voice agents, priced below ElevenLabs on the same per-character unit. | Cartesia's documented advantage is latency and price. On the independent Speko benchmark, sonic-3.5 scores 1574 naturalness Elo—within 16 points of ElevenLabs v3 Conversational—at first-audio p50 latency of 121ms versus 266ms, and at the same $50 per 1M characters. Independent analysis places Sonic 3.6 at roughly $49 per 1M characters against $100 per 1M for Eleven v3, a ~2:1 gap. Cartesia's entry price is also lower (Free, $4/month Pro). Its language coverage is narrower than ElevenLabs' 70+, and its expressive control surface (audio tags, multi-speaker dialogue) is less developed. Cartesia is the stronger choice for real-time agents where time-to-first-audio and cost dominate; ElevenLabs is the stronger choice for expressive offline content and broad language coverage. | Real-time voice agents, conversational IVR, and latency-sensitive applications where sub-150ms time-to-first-audio and lower per-character cost matter more than expressive range or language breadth. | Free tier available. Pro from $4/month. Sonic 3.6 approximately $49 per 1M characters; agent minutes around $0.06. Check the official page for current tiers and concurrency. |
| OpenAI TTS OpenAI TTS guide (official) Independent provider comparison (OpenAI positioned as ecosystem choice) | OpenAI's text-to-speech API, built into the OpenAI platform with a small set of built-in voices, streaming output, and delivery instructions, positioned as an integrated component for teams already using OpenAI models. | OpenAI TTS's advantage is ecosystem integration, not voice quality or control. It offers a limited set of first-party voices, streamed output, and delivery instructions on supported models, and is billed through the same OpenAI account as GPT models. It does not offer Professional Voice Cloning, audio tags, multi-speaker dialogue, or the 70+ language coverage of ElevenLabs v3, and its independent naturalness scores trail both ElevenLabs and Cartesia. The contract surface is simpler: there is no separate Creative/Agents/API family split, no attribution requirement, and no own-voice-only cloning restriction because cloning is not offered. For teams whose product is already built on OpenAI, TTS is a low-friction default that avoids a second vendor; for teams where voice quality, expressive control, or language breadth is the product differentiator, it is underpowered. | Teams already standardized on OpenAI models that need adequate, low-friction TTS inside an existing pipeline and do not require expressive control, voice cloning, or broad multilingual coverage. | Billed per character through the OpenAI API, with rates depending on the model and voice. No separate subscription; no commercial-license gate beyond the API terms. Check the official pricing page for current per-1M-character rates. |
Comparison checked against official product information on 2026-11-04. Features and pricing may change.
Frequently Asked Questions
Can I use ElevenLabs for free commercially?
No. ElevenLabs' Free plan has no commercial license. Content generated on Free and published—whether a YouTube video, client work, a game, or a podcast—cannot be monetized and must credit ElevenLabs with a link to elevenlabs.io. The commercial license starts at Starter ($6/month). This is a rights restriction, not a volume limit: buying extra credits on Free is not an option, and the only way to use Free output commercially is to upgrade before generating the content.
Can I clone someone else's voice with their consent?
Not through ElevenLabs' Professional Voice Cloning pipeline. Professional Voice Cloning requires identity verification and is limited to the user's own voice; even with the other person's explicit consent, cloning someone else is prohibited. Instant Voice Cloning on Starter is intended for your own voice as well. The Voice Library appendix states that user voice models must be based on the user's own voice and that modified versions of one's own voice are also restricted. The only compliant route for third-party voices is licensed talent or the Iconic Marketplace for celebrities, not self-serve cloning. At least 12 US states have voice-cloning consent laws, and ElevenLabs has suspended accounts for violations.
Why does my batch generation stall even when I have credits left?
Because concurrency is a plan-level limit, not a credit limit. Free allows 2 simultaneous Multilingual requests; Starter 3; Creator 5; Pro 10; Scale and Business 15. Exceeding the limit queues or throttles requests rather than billing overage. Pay-as-you-go top-ups raise the character budget but never the concurrency ceiling. For batch dubbing, audiobook production, or any workload that needs parallel generation, the binding constraint is the plan's concurrency column, not the credit balance, and the upgrade decision is often forced by throughput before volume.
Is Eleven v3 suitable for a real-time voice agent?
No. ElevenLabs' own documentation describes v3 as unsuitable for real-time use, and it caps requests at 5,000 characters (roughly five minutes). It is an expressive offline model for narration and dialogue. Real-time agents should use Flash v2.5 (~75ms latency, 40,000-character requests) or v3 Conversational (~280ms). Independent field reports also note that v3's audio tags can occasionally flip a speaker's accent, that tag effectiveness varies by voice and context, and that the overall experience has been described as still feeling like a beta, with ElevenLabs acknowledging ongoing work.
How does ElevenLabs compare with Cartesia on price and latency?
On the independent Speko benchmark, ElevenLabs eleven_v3_conversational scores a naturalness Elo of 1590 with first-audio p50 latency of 266ms; Cartesia sonic-3.5 scores 1574 at 121ms, at the same $50 per 1M characters. Independent analysis also places Cartesia Sonic 3.6 at roughly $49 per 1M characters against $100 per 1M for Eleven v3, a roughly 2:1 gap. ElevenLabs' advantage is concentrated in expressive offline narration and language breadth (70+ languages); for live agents where latency and cost dominate, Cartesia and Flash v2.5 are the appropriate choices, and ElevenLabs' own v3 is not the model to use.
Alternative Tools
-
Suno Suno is an AI music generator with a free tier whose songs are permanently non-commercial and cannot be retroactively licensed, and paid tiers whose 20/60 monthly download caps are far tighter than their credit allowances—with full production workflows requiring Premier and Suno Studio. -
Udio Udio is a high-capacity AI music generator whose downloads have been disabled platform-wide since its October 2025 UMG settlement, whose terms grant no commercial-use rights on any plan, and whose output is owned by Udio or its licensors—making it a song-exploration lab rather than a production or distribution tool. -
Whisper (OpenAI API) Whisper (whisper-1) is OpenAI’s legacy hosted speech-to-text API, priced at $0.006 per audio minute. It is deprecated with a shutdown date of February 26, 2027, has a 25 MiB file ceiling, and does not support streaming. The open-source model remains available for self-hosting. -
Adobe Firefly Adobe Firefly is a generative AI platform built around commercially safe, licensed training data and deep integration with Adobe’s creative applications, offering image, video, audio, and partner-model generation through a credit-based subscription system. -
Bolt.new Bolt.new is StackBlitz's in-browser AI full-stack app builder. It runs a Node.js runtime inside your browser tab, installs packages, starts dev servers, and generates full-stack code from prompts—no local environment setup required. -
Browser Use Browser Use is an MIT-licensed open-source AI browser agent (Python and TypeScript) that lets LLMs control a browser, paired with a usage-based cloud offering at $0.02 per browser-hour for managed stealth browsers, CAPTCHA solving, and a hosted agent API.