Resemble AI in 2026: Detection Pivot and Pricing
Resemble AI in 2026: it stopped selling voice generation to new customers and now focuses on deepfake detection. Its pricing and Chatterbox explained.
Resemble AI no longer sells voice generation to new customers. In an August 13, 2026 post, the company said it is "supporting our existing customers but not taking new ones." It now sells deepfake detection, watermarking and identity tools. Its Chatterbox voice models stay open and free under the MIT license.
What is Resemble AI in 2026?
Resemble AI used to be known for voice cloning and text to speech. Its own pages now describe it as a deepfake detection company. The homepage talks about real-time detection of AI-generated audio, video and images, and it does not mention voice cloning or text to speech.
| Date | What happened | Source | |---|---|---| | 2026-03-27 | Watermarking extended from audio to images and video; Chrome extension beta launched | Resemble blog post | | 2026-08-13 | Voice generation closed to new customers; Chatterbox and DramaBox moved to Voice AI Research | Resemble post | | 2026-10-17 | Homepage positions the company around detection | resemble.ai |
The August post says Resemble has retired "more than 300 older posts on voice cloning." It also says its open models were "downloaded more than 17 million times on Hugging Face." That download figure is a vendor claim, and we did not verify it.
Does Resemble AI still offer voice cloning and text to speech?
Not for new customers. Resemble's API documentation now includes an archive notice. It says Chatterbox and the generative voice models are "no longer offered as commercial paid products," and that its voice API guides "describe the former hosted voice platform." Those guides do not imply that commercial voice services are still available.
Resemble has not published a retirement date for existing customers. Do not assume they lose access, and do not assume they keep it. Ask the company directly if you depend on the service.
Older archived documentation described the former rapid cloning feature as needing "10 seconds - 3 minutes" of audio, and it said the Voice Cloning API required a Business plan or higher. Treat those details as past tense only.
What Resemble sells now
Resemble's current platform is built around five APIs. Each one answers a different question:
- Detect: "Is this real?" Returns a verdict for audio, video or image.
- Watermarker: "Where did this come from?" Embeds and reads watermarks, and is C2PA compatible.
- Identity: "Whose voice or face is this?" Speaker verification, with enrollment from 4 seconds of audio.
- Signal: "Have we seen this attack before?" Matches fraud patterns without transcription or stored content.
- Intelligence: "Why should anyone believe the verdict?" Gives plain-language reasoning and structured JSON.
Resemble's voice AI platform page describes how the watermark works. The PerTh watermark survives resampling, re-encoding, compression, noise, pitch shifting and time-stretching. Detection returns a yes or no answer for whether a watermark is present. The company describes the feature as "EU AI Act Article 50 oriented," which is its own wording.
For context, the European Commission says the AI Act's transparency obligations under Article 50 became applicable on 2 August 2026, and a law-firm summary says the machine-readable marking obligation starts on 2 December 2026. Check the official text before you rely on either date for compliance.
Resemble AI pricing
Resemble's pricing page lists four tiers as of October 2026. Flex is pay as you go, and its credits never expire.
| Plan | Monthly price | Seats | Notes | |---|---|---|---| | Flex | $0, pay as you go | 1 | Credits never expire, no card needed | | Team | $350, or $280 a month billed annually | 5 | Team plan, billed monthly or annually | | Business | $1,000, or $800 a month billed annually | 20 | SSO | | Enterprise | Custom | Custom | SLAs, custom model training, on-premises deployment |
Detection is billed per unit of media. The per-second rates below come from the same pricing page.
| Service | Flex | Team | Business | |---|---|---|---| | Detect audio, per second | $0.035 | $0.015 | $0.015 | | Detect video, per second | $0.07 | $0.03 | $0.03 | | Intelligence, per second | $0.025 | $0.015 | $0.015 |
Multiplying the per-second rate by 60 gives a rough cost per minute of audio. Our arithmetic gives $2.10 per minute on Flex and $0.90 per minute on Team and Business for audio detection. Video detection comes to $4.20 per minute on Flex and $1.80 per minute on Team and Business. Identity search costs $0.0005 per call, and agent detection costs $1 per 1,000 sessions. The page says volume pricing is available.
How accurate is Resemble Detect?
Resemble claims Detect is accurate up to 99.5 percent of the time. It also says Detect was tested against more than 250 generative AI models, returns a verdict in under 300 milliseconds, and covers 54 languages validated against the MLAAD dataset. These are vendor claims. The pages we read name the dataset but show no benchmark results or test dates.
Resemble's March 2026 post reports that the image deepfake detection F1 score rose from 0.961 to 0.970. At MWC Barcelona, it says, "around 70% could not tell the difference" in an audio listening game at its booth. That is a small, unscientific sample run by the vendor, so treat it as a demonstration, not a measurement.
Resemble also reports that at least 15,736 people were victimized across 821 documented attacks in six months, with around 3.46 million files behind them. That is vendor-compiled data, so cite it as Resemble's own figure.
Chatterbox: the open models
Chatterbox is the part of Resemble's work that remains freely available. Its models are listed on the GitHub repository and on Hugging Face under the MIT license. Check the license text for your own use.
| Model | Parameters | Languages | Notes | |---|---|---|---| | Chatterbox-Turbo | 350M | English | Smaller English model | | Chatterbox-Nano | 110M | English | Runs on CPU; the README reports 3x faster than realtime on 8 CPU cores | | Chatterbox (original) | 500M | English | Original English model | | Chatterbox-Multilingual V3 | 500M | 23 languages | Six single-language packs are also listed |
Chatterbox supports zero-shot voice cloning from a reference clip, and it includes an emotion exaggeration control. Every output carries the PerTh watermark. The README does not state a minimum reference length, so test with your own audio. Resemble's "sub 200 ms" figure refers to its former paid service, not to the open models.
Resemble AI alternatives for voice generation
If you want to generate speech, the pricing and models in our PlayHT alternatives guide cover current options. Chatterbox is the free option in that comparison, and it needs your own hardware.
Hume's Octave model is also being shut down, with API access ending on November 13, 2026. Do not choose it as a replacement for Resemble.
Where OndaLabs fits
Readers who came to Resemble for cloning need a place to create custom voices. OndaLabs custom voices clone from a recording you have permission to use, or design a voice from a text description. Each one costs 1,500 credits. Creator, at $22, includes 3 custom voices. Pro, at $49, includes 10, and Business, at $199, includes 30. You can manage custom voices in the app.
Commercial use is included on Starter and above. Detection, watermarking and identity checks are not among OndaLabs' tools, so if you need those, you will need a different product. For plan details, see the pricing page, or read about text to speech to see the generation side.
Frequently asked questions
Is Resemble AI still available?
Yes, but its focus has changed. The company says it is supporting existing customers and is no longer taking new ones for voice generation. Its current product line is deepfake detection, watermarking and identity tools.
Does Resemble AI still offer voice cloning?
Not for new customers, according to the company's August 13, 2026 post. Its archived documentation describes the former hosted voice platform. No retirement date has been published for existing customers.
How much does Resemble AI cost?
Flex is pay as you go at $0 a month, with credits that never expire. Team costs $350 a month, or $280 a month billed annually. Business costs $1,000 a month, or $800 a month billed annually. Enterprise is custom priced.
What is Resemble Detect and how accurate is it?
Resemble Detect is an API that returns a verdict on whether audio, video or images are AI-generated. Resemble claims up to 99.5 percent accuracy and says it was tested against more than 250 models. The pages we read show no benchmark results or test dates, so treat the accuracy figure as a vendor claim.
Is Chatterbox free to use?
Chatterbox is released under the MIT license and can be run on your own hardware. Resemble describes the models as open and free to build on. Read the license text, and check the model card for each specific model before you use it commercially.
What is PerTh watermarking?
PerTh is Resemble's imperceptible watermark, which survives resampling, compression and other edits. Detection returns a yes or no answer for whether the watermark is present. Chatterbox adds it to every output, and Resemble's watermarking API covers audio, images, video and text.