Resemble AI public-data review for creators and small businesses

What changed, in Resemble's own words

Resemble AI spent years as a voice-cloning and text-to-speech vendor, and that is how most directories — including this one until today — still list it. The company's own framing is now blunter than any of them. Its post is filed under the slug why-a-former-voice-ai-company-went-all-in-on-deepfake-detection, and dated 13 August 2026.

The load-bearing quotes:

• "We're not selling voice AI to new customers, we have put the force of the whole company behind detection, for audio, video, and image."
• "We're supporting our existing customers but not taking new ones."
• "The voice models we built are open and free to use, but we won't be producing new ones for commercial sale."

Resemble attributes the shift to demand documented in its own H1 2026 deepfake threat report, which it says recorded at least 15,736 people victimised across 821 documented attacks in six months. We are reporting that figure as the vendor's, not verifying it.

Three independent checks line up with the announcement. Resemble's official pricing page prices only Detect, Intelligence, Identity and Watermarker — there is no TTS, cloning or voice-agent row on it. Its platform page is organised around Detect and Verify, with the voice products demoted to a "Voice AI Research" list in the footer. And its developer documentation still documents the voice APIs, which is consistent with continuing to serve existing customers rather than with selling to new ones.

A note on the TTS prices you will find elsewhere. Several review sites and pricing trackers still quote Resemble TTS at $0.0005 per second, with add-ons of $2–$5 per month per voice clone and $20 per seat. Those numbers are internally consistent and may well be accurate for existing accounts, but no official Resemble page publishes them, and one of them collides exactly with the published Watermarker encode rate of $0.0005 per call. Treat them as unconfirmed.

What Resemble AI costs now

These are the official plans as listed in September 2026. Seats and features are the vendor's; the per-seat arithmetic is ours.

The annual discount is exactly 20% on both paid tiers — $840 a year off Team, $2,400 off Business, as Resemble's page states. Flex is a genuine $0 entry with pay-as-you-go credits and one seat, so there is no minimum commitment to start.

The per-seat column contains the one clean structural finding here: Business costs 28.6% less per seat than Team ($40 against $56 annually, or $50 against $70 monthly), because it charges 2.86x the price for 4x the seats. That inverts the pattern CartSignal has found repeatedly elsewhere in this category, where moving up a tier costs more per unit — HeyGen Business runs about 2x Pro per credit, and Synthesia Creator costs about 17% more per credit than Starter on annual billing. If you are buying seats rather than volume, Resemble's ladder rewards scaling up.

The published usage rates

Usage is billed on top of the plan. Resemble prints each service as a range, from a starting rate down to a volume rate. Per-minute and per-hour columns are our arithmetic.

An hour of continuous audio detection therefore runs $126.00 at the starting rate and $54.00 at the volume rate; an hour of video runs $252.00 and $108.00. Video is priced at exactly 2x audio at both ends of the range, which is the one perfectly consistent relationship in the card.

The discount is not applied evenly

Compare Intelligence against audio detection and the volume curve does something odd. Intelligence starts 28.6% cheaper than audio Detect ($0.025 against $0.035), which makes it look like the budget option. But both land at the same $0.015 at volume. That means the discount is 57.1% on audio detection and only 40% on Intelligence — so the saving that makes Intelligence attractive at low volume disappears completely at high volume, where the two cost exactly the same per second. If you are modelling a large deployment, do not carry the entry-level price gap into the forecast.

Watermarking costs four orders of magnitude less than detecting

This is the number that should drive the buying decision, and the rate card does not spell it out because the two services are billed on different units.

Detection is metered per second. Watermarking is metered per call — a flat $0.0005 to encode and $0.0002 to decode, regardless of how long the file is. Our arithmetic on the starting rates:

Because marking is flat and detecting scales with duration, the gap widens the longer the file gets — and it stays enormous even at the volume rate, where a marked minute still costs 1,800x less than a detected one. Checking a watermark you placed yourself is cheaper again, at $0.0002 a call.

The operational rule falls straight out of it: if you control the pipeline, watermark at creation. Paid detection is for content you did not create and cannot mark — inbound calls, submitted media, monitored broadcast. Budgeting detection against your own output when you could have marked it is the single most expensive mistake available on this rate card.

What the rate card does not tell you

Resemble prints a range but does not publish which plan gets which rate. Flex, Team and Business share one column spanning $0.035 down to $0.015, with volume pricing attached to Enterprise. So you cannot compute your bill, or work out whether a paid plan pays for itself, from public information.

What can be bounded is the best case. If Team's $280/month annual bought the full discount on audio detection — the most generous possible reading — the $0.02 per second saved would repay the subscription at 14,000 seconds, or about 3.9 hours of audio a month. On video, where the saving is $0.04 a second, it would repay at about 1.9 hours. Business at $800/month annual would need roughly 11.1 hours of audio or 5.6 hours of video. Every one of those is a floor: any discount smaller than the full range pushes the break-even higher, and the real figure is not public. Get your tier's actual rate in writing before committing.

The voice models are still free

The part of this story that genuinely helps voice buyers is that Resemble did not take its technology away — it stopped charging for it.

Chatterbox is Resemble's open-source text-to-speech family, released under the MIT licence in three variants: Chatterbox, Chatterbox Turbo (which Resemble calls the "fastest open-source TTS") and Chatterbox Multilingual, supporting "23+ languages" with zero-shot voice cloning. Resemble's terms are unusually plain: you "can use them in commercial products, self-host, modify the weights, and ship to production — no royalties, no revenue share, no usage caps", with no API keys, rate limits or sign-up required to self-host. The code is at github.com/resemble-ai/chatterbox.

The catch is the obvious one: you supply the GPUs, the engineering and the uptime. That is a real cost, just not a licence fee, and it is why Chatterbox is not a like-for-like replacement for a hosted API such as ElevenLabs or Cartesia for most small teams. Resemble has also said it will not produce new commercial voice models, so the ceiling on Chatterbox's quality is whatever its research cadence delivers.

One feature deserves attention from anyone publishing into the EU. Resemble states that every audio file Chatterbox generates carries PerTh, its perceptual-threshold neural watermark, by default. A free, MIT-licensed, self-hostable model that marks its own output is a rare combination against the marking duties discussed below — most hosted vendors we have costed out either do not document machine-readable marking or do not make it available at the entry tier.

Watermarking and detection, as products

PerTh is Resemble's watermarker. The open-source version is MIT-licensed, free and audio-only, and Resemble claims "99.9% decode accuracy" and that the mark "survives MP3 compression, audio editing, noise, and codec transforms", maintaining "~100% detection accuracy even after significant processing". PerTh Multimodal extends the same idea to audio, video, image and text with real-time encode and decode via API, and is Enterprise-only — worth knowing if multimodal marking is the reason you are looking at Resemble at all, because it is not available on Flex, Team or Business.

Detect is the detection side, and its public pages carry three model generations at once: DETECT-2B, DETECT-3B Omni, and DETECT-World, which Resemble calls "the first world model for deepfake detection". Published accuracy claims are per-generator — "up to 99.5%" overall, with ">99%" against StyleGAN and Veo, 99% against Flux, Gemini and GPT-4o, 98% against DALL·E 3 and Midjourney, and 94% against Stable Diffusion.

Two caveats on those numbers. First, the language counts do not agree across Resemble's own pages: the Detect product page says "54 Languages covered, validated against MLAADv10", while its DETECT-2B announcement describes 94–98% precision across "30+ languages". Both may be true of different models, but nothing public maps which figure belongs to which. Second, and to Resemble's credit, MLAADv10 is a named public benchmark — that is better disclosure than most accuracy claims in this category, including several CartSignal has declined to endorse elsewhere. Resemble's compliance page separately cites DETECT-3B Omni at "97%+ audio accuracy" with "<300ms" latency.

A correction to Resemble's own EU AI Act page

Resemble markets heavily into EU AI Act Article 50 labeling duties, and its compliance page states the penalty as "€35M fine per violation or 7% global revenue".

That is the wrong tier for the obligation it is selling against. Under Article 99 of the AI Act, the €35M / 7%-of-turnover ceiling attaches to breaches of the Article 5 prohibited practices. Transparency obligations of the kind set out in Article 50 sit in the lower tier — up to €15M or 3% of total worldwide annual turnover, whichever is higher — which is the figure CartSignal's own Article 50 explainer uses. Resemble's page cites no article for its number.

The duty itself is real, it took effect on 2 August 2026, and marking your output is genuinely the cheap way to meet it. But buy against the correct exposure, not a figure inflated by more than 2x.

Who it is for

It is no longer for voice buyers. If you came here for hosted text-to-speech or voice cloning, Resemble will not sell it to you. On published list prices the alternatives CartSignal has costed out are Cartesia at roughly $0.038 per TTS minute, ElevenLabs at about $0.18 a minute on Starter with the broadest feature set, and Murf AI for timeline-based studio voiceover with team seats. The full shortlist is in ElevenLabs alternatives and the head-to-head in ElevenLabs vs PlayHT. Self-hosting Chatterbox is the option only if you have the infrastructure to run it.

It is for teams with a detection or provenance problem. Trust and safety, fraud prevention on inbound voice, broadcast and platform monitoring, and EU compliance programmes that need machine-readable marking. The Flex tier at $0 with pay-as-you-go credits makes it cheap to trial honestly, and the watermarking rates are low enough that marking your entire output is close to free.

It is worth noting how unusual this listing now is. PlayHT is on this site as a discontinued product with a dead domain. Resemble is the opposite case: a healthy company that deliberately walked away from the category the directory listed it in. Both are here because older guides still recommend them for something you can no longer buy.

Frequently asked questions

Can you still buy voice cloning or text-to-speech from Resemble AI?

Not as a new customer. Resemble wrote on 13 August 2026 that it is "not selling voice AI to new customers" and is "supporting our existing customers but not taking new ones". Its official pricing page carried no TTS, cloning or voice-agent price when we checked on 9 September 2026 — only detection, Intelligence, Identity and Watermarker rates. The per-second TTS figures still circulating on review sites have no official page behind them.

How much does Resemble AI cost in 2026?

Flex is $0/month with 1 seat, Team is $350/month or $280/month annual with 5 seats, Business is $1,000/month or $800/month annual with 20 seats, and Enterprise is custom — a flat 20% annual discount on both paid tiers. Usage is extra: audio detection from $0.035/second, video from $0.070/second, images from $0.035 each, Intelligence from $0.025/second, Identity search $0.0005/call, Watermarker encode $0.0005 and decode $0.0002 per call.

Is watermarking cheaper than deepfake detection?

By four orders of magnitude, because they bill on different units. Encoding a watermark is a flat $0.0005 per call whatever the file length; detection is $0.035 per second of audio. Marking one minute costs $0.0005 against $2.10 to detect it — about 4,200x — and because marking is flat the ratio grows to roughly 252,000x on a one-hour file. If you control the pipeline, watermark; pay for detection only on content you did not create.

Is Resemble AI's voice technology still available for free?

Yes, as Chatterbox under the MIT licence, in Chatterbox, Turbo and Multilingual variants, the last covering "23+ languages" with zero-shot cloning. Resemble says you can ship it commercially with "no royalties, no revenue share, no usage caps", self-hosted with no sign-up. Every file it generates carries Resemble's PerTh watermark by default. You supply the infrastructure, and Resemble says no new commercial voice models are coming.

Does Resemble AI overstate the EU AI Act penalty?

Its compliance page cites "€35M per violation or 7% global revenue" against Article 50-style marking duties. Under Article 99, that ceiling applies to Article 5 prohibited practices; Article 50 transparency breaches sit in the lower tier of up to €15M or 3% of worldwide annual turnover. Resemble's page names no article for its figure. The obligation is real — the exposure is smaller than stated.

Source links

• Official blog, 13 August 2026: why Resemble stopped selling voice AI
• Official pricing page: plans, seats and the usage rate card
• Official platform page: Detect and Verify positioning
• Official Detect product page: model names, accuracy and MLAADv10
• Official Watermarker product page: PerTh and PerTh Multimodal
• Official Chatterbox model page: MIT licence and variants
• Official Chatterbox repository
• Official EU AI Act page, source of the €35M / 7% claim
• Official developer documentation