ElevenLabs can generate speech in a cloned human voice from a few minutes of audio, so anyone buying, building on or defending against it needs to know exactly which models, plans and safeguards apply. ElevenLabs is an AI audio platform from the company of the same name, offering text to speech, voice cloning, voice design, speech to text, sound effects, music generation and real-time voice agents through a web app and an API. Its current flagship speech model, Eleven v4, was released on 28 September 2026, following Eleven v3, which launched as an alpha research preview on 3 June 2025. This page sets out the model line-up with exact API identifiers, the plans and what each unlocks, the two voice cloning tiers and their verification rules, provenance and labelling, data handling, and the release history. It is written for creators, product and procurement teams, developers, and the security and fraud teams who deal with synthetic voice.
Last checked: 2026-11-15.
This is an independent guide from NHI Mgmt Group, based on ElevenLabs' public documentation as of November 2026. It is not sponsored or reviewed by ElevenLabs. Product features change, so confirm settings against ElevenLabs' current documentation.
Key takeaways
- Eleven v4 is the current flagship. eleven_v4 and the real-time eleven_v4_turbo were released on 28 September 2026 and support 90+ languages, according to the ElevenLabs models page and changelog.
- Two cloning tiers with different checks. Instant Voice Cloning starts on the Starter plan from about a minute of audio; Professional Voice Cloning starts on Creator, needs at least 30 minutes of audio and a spoken verification that the voice is your own.
- Commercial use starts at Starter. The pricing page lists a commercial licence and music commercial use from the $6 a month Starter plan; the Free plan does not include them.
- Provenance is layered. ElevenLabs states that it incorporates C2PA, runs a public AI Speech Classifier, and is among the companies Google named in May 2026 as bringing SynthID watermarking to their content.
- Training use is opt-out. The privacy policy says personal information, including inputs and outputs, may be used to train models unless you opt out under Data use; Enterprise customers can request Zero Retention Mode for eligible API calls.
What is ElevenLabs?
ElevenLabs is a generative audio provider. Its products cover the main synthetic audio tasks: turning text into speech in a stock, designed or cloned voice; changing one voice into another (speech to speech); transcribing audio (the Scribe models); generating sound effects and music from text; dubbing; and running conversational voice agents (ElevenAgents) on the Eleven Realtime models. Access is through the web app, which bundles these as studio tools, and the developer API, which exposes each model by ID.
For how ElevenLabs sits among other image, video, voice and music models, see the AI Image, Video, Voice and Music Models Hub. For conversational voice models from several providers side by side, see Realtime Voice AI Models Compared.
Key facts
| Field | Value (as of 2026-11-15) |
|---|---|
| Provider | ElevenLabs |
| Modalities | Speech synthesis, voice cloning and design, speech to speech, speech to text, sound effects, music, voice agents |
| Flagship speech model | eleven_v4 (released 28 September 2026) |
| Real-time speech model | eleven_v4_turbo, about 100 ms median latency |
| Languages (v4) | 90+ |
| Access | Web app, API, ElevenAgents |
| Weights | Closed (reached through app or API) |
| Plans | Free, Starter, Creator, Pro, Scale, Business, Enterprise |
| Provenance | C2PA (per ElevenLabs safety page), AI Speech Classifier, SynthID adoption announced by Google |
Models and versions
ElevenLabs publishes a model ID for every capability. The table lists the active models on the models page, with the language count, the per-request character limit and the median latency where the page states them. Latency figures are medians that exclude application and network latency.
| Model ID | Purpose | Languages | Character limit | Median latency |
|---|---|---|---|---|
| eleven_v4 | Expressive text to speech, top tier | 90+ | 10,000 (about 10 min) | Not stated by ElevenLabs |
| eleven_v4_turbo | Expressive real-time text to speech | 90+ | Not stated by ElevenLabs | About 100 ms |
| eleven_v3 | Expressive text to speech | 70+ | 5,000 (about 5 min) | Not stated by ElevenLabs |
| eleven_v3_conversational | Expressive real-time text to speech | 70+ | Not stated by ElevenLabs | About 165 ms |
| eleven_multilingual_v2 | Lifelike text to speech | 29 | 10,000 | Not stated by ElevenLabs |
| eleven_flash_v2_5 | Low-latency text to speech | 32 | 40,000 (about 40 min) | About 100 ms |
| eleven_flash_v2 | Low-latency text to speech, English | 1 | 30,000 | About 135 ms |
| eleven_ttv_v3, eleven_multilingual_ttv_v2 | Voice design (text to voice) | 70+ / 29 | Not stated by ElevenLabs | Not stated by ElevenLabs |
| eleven_multilingual_sts_v2, eleven_english_sts_v2 | Speech to speech voice changer | 29 / 1 | Not stated / 10,000 | Not stated by ElevenLabs |
| eleven_text_to_sound_v3 | Sound effects from text | Not applicable | Not stated by ElevenLabs | Not stated by ElevenLabs |
| scribe_v3, scribe_v3_live | Speech to text, recorded and real time | 90+ | Not applicable | Not stated by ElevenLabs |
| scribe_v2_medical | Speech to text for clinical audio | 90+ | Not applicable | Not stated by ElevenLabs |
| eleven_realtime_v1_mini, eleven_realtime_v1_max | Voice agent models (fast, complex multi-step) | 90+ | Not applicable | Not stated by ElevenLabs |
| music_v2_5, music_v2 | Music generation | "en, es, de, ja, and more" | Not applicable | Not stated by ElevenLabs |
As of 2026-11-15, from the ElevenLabs models page. Eleven v3 supports inline audio tags in lowercase square brackets, such as [whispers] or [laughs], to direct delivery, as described in the Eleven v3 launch post.
Deprecated models and removal dates
| Deprecated model | Replacement named by ElevenLabs | Removal date |
|---|---|---|
| eleven_turbo_v2_5 | eleven_flash_v2_5 (functionally equivalent, lower average latency) | Not stated by ElevenLabs |
| eleven_turbo_v2 | eleven_flash_v2 | Not stated by ElevenLabs |
| scribe_v1 | scribe_v2 (itself succeeded by scribe_v3) | Not stated by ElevenLabs |
| music_v1 | music_v2_5 | 31 January 2027 |
| eleven_text_to_sound_v2 | eleven_text_to_sound_v3 | 1 March 2027 |
Developers pinning model IDs in code or agent configuration should search for these strings before the removal dates. The wider pattern of media model retirements is tracked in the AI Media Model Retirements Tracker.
Pricing and plans
ElevenLabs sells monthly credit allowances. Credits are spent across the products in the plan. The table shows the recurring monthly price; on the check date the pricing page also showed an introductory first-month price of $1 on Starter and $11 on Creator. Prices exclude taxes.
| Plan | Price a month | Credits a month | Cloning | Other listed features |
|---|---|---|---|---|
| Free | $0 | 10,000 | Not listed on Free | Core products, no commercial licence |
| Starter | $6 | 30,000 | Instant Voice Cloning | Commercial licence, music commercial use |
| Creator | $22 | 121,000 | Adds Professional Voice Cloning (1 slot) | 128 and 192 kbps audio options |
| Pro | $99 | 600,000 | As Creator (1 PVC slot) | 44.1 kHz PCM output via API, 192 kbps audio |
| Scale | $299 | 1,800,000 | 3 Professional Voice Clones | 3 seats |
| Business | $990 | 6,000,000 | 10 Professional Voice Clones | 10 seats, low-latency TTS "as low as 5c/minute" |
| Enterprise | Custom | Custom | More voices (no count given) | Custom terms around DPA and SLAs, BAAs for HIPAA customers |
As of 2026-11-15, from elevenlabs.io/pricing. API usage pricing is published separately by ElevenLabs. If you downgrade below Creator, a Professional Voice Clone stays on the account but cannot be used until you upgrade again, according to the cloning documentation.
Features
There are two tiers, and the difference matters for both quality and abuse risk.
| Field | Instant Voice Cloning (IVC) | Professional Voice Cloning (PVC) |
|---|---|---|
| Lowest plan | Starter | Creator |
| Audio needed | 1 to 2 minutes recommended; each file at least 5 seconds, total at least 10 seconds | Minimum 30 minutes; 2 to 3 hours recommended |
| Method | Clone built from short samples | Fine-tuning request on the samples; PVC fine-tuning on Eleven v4 available to every workspace from 2 November 2026 |
| Ownership rule | Governed by the Terms of Service and Prohibited Use Policy | "You can only create a Professional Voice Clone of your own voice" |
| Verification | No separate ownership step stated in the cloning guide | Spoken verification: read lines aloud, matching the training set-up; retry after 24 hours if it fails |
| Zero Retention Mode | Not eligible | Not eligible |
As of 2026-11-15. The ElevenLabs safety page adds that the platform blocks cloning of celebrity and other high-risk voices, and the cloning guide states that all audio generated by its models can be traced back to the user responsible for the generation. ElevenLabs recommends MP3 at 192 kbps or above for cloning samples.
For defenders, the practical point is that a usable instant clone needs only about a minute of clean speech, which is easily taken from a podcast, webinar or voicemail greeting. That is why voice alone should not authorise payments or account changes. See the Voice Cloning and Family Emergency Scams Guide and the voice biometrics and voiceprint authentication guide.
Availability
ElevenLabs is reached through its web app, which bundles text to speech, voice cloning, dubbing, sound effects and music as studio tools; through the developer API, which exposes each model by ID; and through ElevenAgents for conversational voice agents. Plans run from Free through Starter, Creator, Pro, Scale and Business to Enterprise, and features such as the commercial licence and voice cloning tiers depend on the plan. Country availability is not covered on the pages checked for this guide.
Provenance and labelling
Three mechanisms are documented, and they work differently.
| Mechanism | What it does | Source and limits |
|---|---|---|
| C2PA | Open standard for signed provenance metadata | ElevenLabs safety page says it incorporates C2PA; metadata can be lost when files are re-encoded |
| AI Speech Classifier | Public tool returning the probability that a clip was generated with ElevenLabs | Uses the first minute only, no login; does not reliably classify Eleven v3 audio; checks for ElevenLabs output only |
| SynthID | Imperceptible watermark from Google DeepMind | Google's I/O 2026 post (20 May 2026) names ElevenLabs among companies bringing SynthID to their content; rollout detail not stated by ElevenLabs |
As of 2026-11-15. Treat each as evidence, not proof: a negative result from any one of them does not show a clip is human. For the differences between signed metadata and watermarks, see SynthID vs C2PA Content Credentials, and for the cross-product picture see the AI-Generated Media Labelling Tracker. If you publish synthetic voice in the EU, the disclosure duties are covered in the EU AI Act Article 50 Transparency Obligations Guide.
What safeguards does ElevenLabs apply?
- No-go voices: cloning of celebrity and other high-risk voices is blocked, per the safety page.
- PVC verification: the voice owner must pass a spoken verification before a Professional Voice Clone is fine-tuned.
- Moderation: automated classifiers, human review and internal investigations, with enforcement up to bans for persistent or serious cases.
- Usage policy: the Prohibited Use Policy covers inputs (including voice samples) and outputs, treats an attempted violation as a violation, and allows ElevenLabs to require prior approval or extra safeguards for heightened-risk uses. Enforcement ranges from removal to termination, with an appeal route.
- Reporting: anyone who believes their voice was cloned without consent can use the "Unauthorized Use of Voice" report form linked from the policy.
- Traceability: generated audio is traceable to the generating account, which supports investigations and takedown requests.
Data and privacy
| Topic | What ElevenLabs states (as of 2026-11-15) |
|---|---|
| Training use | Personal information, which can include inputs, outputs and voice data, is used to "research, build, train and improve our AI models" |
| Opt-out | In account settings under Data use; applies only to training after the opt-out |
| Voice data retention | Data generated about your voice kept no longer than 3 years after your last interaction, unless law requires or allows otherwise |
| Location | Personal information transferred to the United States for storage; EEA, UK and Swiss transfers rely on adequacy decisions (including the EU-U.S. Data Privacy Framework), Standard Contractual Clauses or derogations |
| EEA controller | ElevenLabs Poland sp. z o.o. |
| Business customers | Processing for business customers is governed by their contracts, not the consumer privacy policy |
| Zero Retention Mode | Enterprise option for select customers, mainly healthcare and banking; set enable_logging=false per request |
Zero Retention Mode scope
When enabled, logging of inputs and outputs of eligible API requests is restricted, along with logging of the account email. Eligible: all /v1/text-to-speech/ endpoints and the TTS WebSocket, /v1/text-to-dialogue/, /v1/speech-to-speech/, /v1/speech-to-text/, the music generation and planning endpoints, POST /v1/sound-generation, and ElevenAgents API traffic when enabled per agent under Privacy, Advanced. Not covered: web UI traffic, support tickets and separately shared files, Instant and Professional Voice Cloning, Dubbing, Music Finetunes, Image and Video, and ElevenCreative Studio. MCP support is unavailable for ZRM and HIPAA users, and usage analytics still record that a request was made.
API keys for ElevenLabs are billing and data-access credentials like any other model provider key. Scope them, keep them in a secrets manager, and rotate on exposure: see the response steps for AI model provider API keys.
Where identity and security fit
Cloned voices undermine any process that treats a familiar voice as proof of identity: helpdesk password resets, payment approvals by phone, and voiceprint authentication. The controls are procedural and identity-based rather than detection-based: call-back to a number on file, out-of-band approval, and phishing-resistant authentication for staff. Voice agents built on ElevenAgents also hold credentials for the systems they call, so they need the same scoping and monitoring as any other non-human identity. See the Voice AI Agent Security and Fraud Guide and the Spotting Deepfake Scams Guide.
Release history
| Date | Event |
|---|---|
| 3 June 2025 | Eleven v3 launched as an alpha research preview, 70+ languages (later made generally available; GA date not stated by ElevenLabs) |
| 28 September 2026 | Eleven v4 and Eleven v4 Turbo released |
| 19 October 2026 | Eleven Realtime model family for ElevenAgents begins gradual rollout (v1 Mini default, v1 Max) |
| 26 October 2026 | Music v2.5 becomes the default for the Studio video-to-music workflow |
| 2 November 2026 | Eleven Realtime available to every workspace; PVC fine-tuning on Eleven v4 available to every workspace |
| 9 November 2026 | Scribe v3 and Scribe v3 Live available in the Speech to Text API; agents falling back from Eleven v3 Conversational use Eleven v4 Turbo or Eleven v4 |
| 31 January 2027 | Scheduled removal of music_v1 |
| 1 March 2027 | Scheduled removal of eleven_text_to_sound_v2 |
As of 2026-11-15, from the ElevenLabs changelog, models page and Eleven v3 launch post.
Limits
- Per-request length: 5,000 characters on eleven_v3 against 10,000 on eleven_v4 and 40,000 on eleven_flash_v2_5; long scripts need chunking on v3.
- Detection gaps: the AI Speech Classifier does not reliably classify Eleven v3 audio and looks only at the first minute.
- Cloning slots: PVC slots are counted per plan (1 on Creator and Pro, 3 on Scale, 10 on Business).
- Zero retention exclusions: cloning and dubbing workflows are outside ZRM, so voice samples for clones follow the standard retention terms.
- Model removals: two deprecated models have dated removals in 2027; the turbo v2 models have named replacements but no published removal date.
Practitioner checklist
- Record which ElevenLabs model IDs your code and agents pin, and plan migrations for music_v1 (31 January 2027) and eleven_text_to_sound_v2 (1 March 2027).
- Confirm the plan covers commercial use (Starter or above) before publishing generated speech or music.
- Obtain written consent from any person whose voice you clone, and keep it with the voice record.
- Decide on training use: set the Data use opt-out for accounts handling personal data, and request Zero Retention Mode where eligible.
- Check that cloning, dubbing and web UI work, which fall outside ZRM, are not used for sensitive audio.
- Store ElevenLabs API keys in a secrets manager, separate keys per environment, and monitor usage for spikes.
- Keep provenance intact: avoid re-encoding steps that strip metadata, and label synthetic voice where law or policy requires.
- Remove voice as a sole authentication factor in helpdesk and payment processes.
Standards and references
- ElevenLabs: Models
- ElevenLabs: Pricing
- ElevenLabs: Changelog
- ElevenLabs: Professional Voice Cloning
- ElevenLabs: Safety
- ElevenLabs: Privacy Policy
- ElevenLabs: Zero Retention Mode
- Google: Google I/O 2026 announcements
Related NHI Mgmt Group resources: AI Image, Video, Voice and Music Models Hub · Realtime Voice AI Models Compared · Voice AI Agent Security and Fraud Guide · SynthID vs C2PA Content Credentials