ElevenLabs Content Provenance API: Power-User Guide to Watermark Checks and STT Prices

Last updated: November 18, 2026

The ElevenLabs Content Provenance API is a beta endpoint, POST /v1/content-provenance/watermark, released on November 17, 2026. It checks an uploaded audio file for the inaudible ElevenLabs watermark and, when it can, returns the product, model, generation time and ID. It never returns the prompt or the account behind the audio.

Contents

What changed on November 17

The ElevenLabs changelog entry for November 17, 2026 lists four headline changes:

  • Content Provenance API (beta): a watermark check for uploaded audio. API keys need the new watermark permission.
  • Lower speech to text prices: $0.20 per hour for batch and $0.22 per hour for realtime transcription on every Scribe model.
  • Tools as code: a tool builder in the Python and JavaScript SDKs, from v2.79.0.
  • Hosted MCP: 39 new tools for building ElevenAgents from any MCP client.

The same entry also deprecates gpt-4.1, gpt-4.1-mini, gpt-4o and gpt-4o-mini in ElevenAgents ahead of their shutdown on January 14, 2027. Requests fall back to gpt-6-luna, for 10% of requests from December 15, 2026 and for all requests from January 9, 2027.

Content Provenance API limits

According to the Content Provenance API documentation, ElevenLabs audio carries an inaudible watermark embedded in the audio samples, and the endpoint reads the file you upload to look for it. These are the published limits:

Item Limit or rule
Endpoint POST https://api.elevenlabs.io/v1/content-provenance/watermark (beta)
Authentication API key with the watermark permission; OAuth access tokens are not accepted
File types Any audio type the decoder can read, including MP3, WAV, FLAC, AAC and Ogg
File size Up to 50 MB, one file per request
Audio length 0.5 seconds to 5 minutes after decoding
Rate limit 25 checks per API key user per 24 hours; more returns 429
Decoder version Optional version field; only latest is supported

An image or other non-audio file returns 400 with the status invalid_file_type. ElevenLabs says a 429 is worth retrying later, while a 400 or 422 means the request itself needs to change.

How to check an audio file

The steps below follow the official documentation:

  1. Create or edit an ElevenLabs API key and grant it the new watermark permission.
  2. Confirm the file is audio, no larger than 50 MB and between 0.5 seconds and 5 minutes long.
  3. Send it as a multipart form with the field name file and the xi-api-key header, for example: curl -X POST https://api.elevenlabs.io/v1/content-provenance/watermark -H "xi-api-key: $ELEVENLABS_API_KEY" -F "file=@./example.mp3;type=audio/mpeg"
  4. Optionally add -F "version=latest"; any other value is rejected.
  5. Read is_detected first, then details if it is present.
  6. Track your usage so you stay under 25 checks per 24 hours, and retry 429 responses later.

Reading the result

A finished check returns two fields. is_detected is true when the decoder recognized an ElevenLabs watermark and false when it did not.

details is filled in only when generation details can be recovered, which might not be possible for short or heavily changed audio. It then carries product, model, generated_at and unique_id, and any of the four can be null. The documentation’s example shows the product “Music” and the model “Eleven Music v2”.

ElevenLabs adds a caution worth repeating: the endpoint only reads the watermark already in the file. For any decision that depends on where audio came from, keep a person in the loop.

Lower speech to text prices

The changelog says the new rates apply to every Scribe model on all self-serve and Business plans; Enterprise pricing is unchanged. The API pricing page lists these rates:

Model Type Price per hour of audio
Scribe v3 Batch $0.20
Scribe v3 Live Realtime $0.22
Scribe v2 Batch $0.20
Scribe v2 Medical Batch $0.20

For model details, see our ElevenLabs Scribe v3 guide. If you only need occasional transcription inside a chatbot, compare our answers on whether ChatGPT can transcribe audio and whether Gemini can transcribe audio.

Hosted MCP and tools as code

Hosted MCP. The November 17 changelog adds 39 tools covering workflow edits, attaching and detaching tools, test runs, agent generation, phone number import, batch calling, workspace secrets and ElevenAgents settings. It says edits are staged in a draft and only the publish tool makes them live. Connected clients ask for consent again because of two new read scopes.

That builds on the November 9 update, which added 24 tools for inspecting agents, covering analytics, insights, alerts and test summaries under the existing convai_read and convai_write permissions. The server lives at https://api.elevenlabs.io/v1/mcp and uses OAuth. Note that the hosted MCP documentation still warns that updates through the server publish directly, so test on a branch before relying on draft behavior.

Tools as code. The tools as code guide explains that you can define webhook, client and system tools with chained calls such as get(), header() and assign(), then pass build() to tools.create() or tools.update(). In Python the builder comes from elevenlabs.agent_tools; in JavaScript it is exported from @elevenlabs/elevenlabs-js. It calls no new API endpoints, and built tools stay editable in the dashboard.

How it compares with other watermark checks

Anthropic took a related step one day earlier. Its release notes say text from Claude Opus 4.6, Claude Opus 4.5, Claude Sonnet 4.5 and Claude Haiku 4.5 carries a text watermark on supported platforms, and that non-default temperature, top_p or top_k values might remove it. Our Claude Text Watermark guide covers that system.

Question ElevenLabs Content Provenance API Claude text watermark
Media Audio Text
When the mark can fail Details may not be recoverable for short or heavily changed audio Non-default sampling settings might leave text unmarked

For another vendor’s approach, see our Grok API watermarking guide.

Frequently asked questions

Is the ElevenLabs Content Provenance API free?

ElevenLabs’ documentation and changelog do not list a separate price for the beta endpoint. Each API key user can submit 25 checks per 24 hours, and ElevenLabs asks you to contact support if you need a higher limit.

Can it detect audio from other AI voice generators?

The documentation describes a check for the watermark ElevenLabs embeds in its own generated audio, so it is not a general AI voice detector. A false is_detected result means no ElevenLabs watermark was found in the file.

Does the API reveal who created a piece of audio?

No. When a watermark matches a generation record, the response can include the product, model, generation time and a unique ID. It never includes the prompt, the account or the user behind the generation.

How much does ElevenLabs speech to text cost after the November 17 update?

Every Scribe model now costs $0.20 per hour of audio for batch transcription and $0.22 per hour for realtime transcription on self-serve and Business plans. Enterprise pricing is unchanged.

Which audio files can I check?

Any audio type the decoder can read, including MP3, WAV, FLAC, AAC and Ogg. Each request takes one file of up to 50 MB that runs between 0.5 seconds and 5 minutes after decoding.

Sources

Image: ElevenLabs logo, public domain (trademarked), via Wikimedia Commons.

Related answers

Leave a Comment