Detection is easy to build and hard to sell, which is why the money is somewhere else. When an Anthropic engineer confirmed on August 11, 2026 that the company will ship "a text detection API that you can use yourself," the first substantive public reply was not about accuracy. It was: so we pay a different API to check if text has a watermark?
That reaction is the whole business question in one line. Claude now watermarks all text from models launched on or after August 2, 2026. A detection service is coming. Somebody is going to have to decide what it costs — and every available answer has a problem.
TL;DR — the monetisation map
| Question | Direct answer |
|---|---|
| Is detection announced? | Yes — a text detection API, per Anthropic. No pricing published |
| Can it be free? | Technically yes, but free = an evasion oracle |
| Can it be paid? | Yes, but paid = blocked independent verification |
| Who actually pays? | Platforms, publishers, enterprises, education — not individuals |
| Biggest revenue pool? | Compliance tooling, not per-check detection |
| Best unpriced value? | Distillation detection — worth a lot, sold to nobody |
| Is C2PA a comparable? | Yes, and it is free — a hard price anchor |
| Counter-market? | Watermark stripping and laundering, immediately |
| Premium watermark-free tier? | Not legally viable for a regulated provider |
The pricing paradox
Start with why this is genuinely hard rather than just unresolved.
Free and unlimited creates an evasion gradient. A detector that anyone can query without limit is also a scoring function for defeating the detector. Paste, check, rewrite, re-check, repeat until the confidence drops below threshold. One critic put it bluntly within hours of the announcement: giving away a way to validate whether the watermark exists is watermarking for people who do not try. That is overstated — the mark still catches everyone who does not bother — but the mechanism he describes is real, and it is why adversarial-ML practitioners have argued against public verification endpoints for a decade.
Paid or gated blocks the accountability the regulation exists for. The EU AI Act's Article 50(2) Code of Practice, which Anthropic signed and which drove this whole rollout, requires supporting third-party detection. If the only party who can check is the party who made the mark, nobody can independently measure the false-positive rate — and a false-positive rate on a signal used in academic misconduct or employment decisions is not a detail. Charging for it puts a meter on transparency.
Rate-limited free tiers are the obvious compromise, and they only partly work. They preserve casual verification and slow bulk iteration, but a determined evader distributes queries. This is where most providers will land anyway, because it is the least-bad option, not because it resolves anything.
There is a fourth option worth naming: identity-gated free access. Free but authenticated, with logging, so bulk iterative probing is visible even when it is not blocked. That trades user privacy for abuse resistance, which is its own uncomfortable exchange on a transparency product.
The free anchor nobody can escape
Any pricing conversation runs into an awkward comparable: C2PA verification is already free.
The open provenance standard behind LinkedIn's Content Credentials rollout, and behind the file-metadata half of Anthropic's own announcement, is checkable today by anyone at zero cost, via multiple independent verifiers. As we documented in how to detect a Claude watermark, the file-provenance half of the system is mature, open, and unpriced.
That sets an anchor. When half of your provenance product is a free open standard and the other half is a proprietary paid API, the pricing looks less like a service and more like a toll on the part you happen to control. The counter-argument is legitimate — statistical text detection genuinely requires the secret key and the provider's infrastructure, while C2PA verification is pure cryptography anyone can run — but anchoring does not care about technical justification.
Where the money actually is
Detection-as-a-product is a thin market. Provenance as an obligation is a thick one. Four revenue pools, roughly in ascending order of size:
1. Per-check API calls — smallest
Individuals checking one document will not sustain a business. Willingness to pay for a single verification is near zero, and the people most motivated to check are typically the least able to pay. This tier exists for completeness and goodwill, not revenue.
2. Platform integration — real but concentrated
Platforms screening uploads at volume have both the need and the budget. Spotify demoting AI-generated artist profiles is exactly this shape of customer: the provenance signal becomes a ranking input, which means it has to be checked continuously at scale. So does every marketplace, journal, job board, and social platform now writing an AI policy.
The catch is that there are perhaps a few dozen such buyers globally, each with enough leverage to negotiate hard, and each capable of building competing heuristics if the price is wrong. Concentrated demand is not the same as strong pricing power for the seller.
3. Compliance and audit tooling — largest
This is where prior regulatory waves put the money, every time. The technical check is a commodity; the attestation, the audit trail, the reporting dashboard, and the "we can demonstrate we complied" artifact are what enterprises actually buy.
An organisation subject to Article 50 does not want a detection endpoint. It wants a defensible record of what it screened, when, under what policy, with what outcome — the same thing SOC 2 tooling sells, packaged for content provenance. That is a recurring-contract business with switching costs, not a per-call one. Our EU AI Act and US policy guide covers the obligations this attaches to.
4. Distillation defence — most valuable, sold to nobody
Here is the pool that never shows up on a price list. A critic in the launch reaction accused Anthropic of building this to protect against distillation rather than for public transparency. Set aside the accusatory framing: detecting whether a rival model was trained on your outputs is worth an enormous amount to a frontier lab, and a watermark is a good instrument for it.
That value is entirely internal. It shows up as litigation leverage, competitive intelligence, and terms-of-service enforcement — not revenue. Which is why it also functions as a strategic subsidy: the lab has a reason to build detection well even if the external product never makes money, and users benefit from infrastructure they are not funding. It is the most honest answer to "why build this at all" and the one least likely to appear in marketing.
The counter-market arrives immediately
Every provenance system creates its own removal industry, and this one is already forming in public. The prior art is unambiguous: education-focused AI detection spawned a matching "humanizer" market within months, and those tools now outnumber the detectors.
For text watermarks specifically, the removal paths are known and cheap:
| Method | Effectiveness | Cost |
|---|---|---|
| Round-trip through a second model | High — destroys statistical marks | One extra inference call |
| Heavy manual editing | High, and the only method that also improves the work | Human time |
| Translation and back-translation | High | Trivial |
| Using an open-weight model instead | Total — no mark ever exists | Free, and covered in our analysis of why open weights break watermarking |
| Stripping Unicode or metadata | Irrelevant for text marks — they are not metadata | Trivial |
The economics here are bleak for the detection seller. Evasion costs one extra API call. Detection costs a subscription. When the counter-measure is an order of magnitude cheaper than the measure, the measure cannot be priced aggressively — and as one commenter noted with some accuracy, you end up paying tokens to add the mark and tokens to remove it, with the provider collecting on both.
The premium tier that cannot exist
A joke circulated widely in the developer reaction: surely the next step is a higher-priced plan that removes the watermark.
It is worth answering seriously, because it clarifies the constraint. No regulated provider can sell an exemption from a compliance mechanism. If Article 50 requires machine-readable marking of AI output, a paid unmarked tier is not a product decision, it is non-compliance with a price attached. The commercial ceiling on watermarking is set by regulation, not by willingness to pay.
The nearest legal cousin is enterprise agreements with contractual assurances about data handling and detection access — selling control over the provenance process rather than exemption from it. That is a real product. It is just much less interesting than the joke.
What this means if you are building in the space
- Do not build a detection API business. The technical check trends to commodity, the price anchor is free, and the primary key-holders are the labs themselves.
- Build the compliance layer instead. Policy management, audit trails, evidence retention, multi-provider aggregation. That is where obligation converts to recurring revenue.
- Assume multi-provider from day one. Google, Anthropic, and whichever labs follow will each have separate schemes and separate keys. Nobody wants four integrations, and the aggregation layer is a defensible position.
- Price on liability, not on volume. Customers with regulatory exposure pay for reduced risk, not for API calls. The buyer is a compliance officer, not an engineer.
- Expect the ground to move. Detection availability, pricing, and even the underlying schemes are unsettled — Anthropic's API is announced but unshipped, and the first published false-positive rate will reset every assumption in this post.
Bottom line
Watermark detection is monetisable, but not as detection. Sold directly it faces a free open-standard anchor, an evasion counter-market that undercuts it by an order of magnitude, and a pricing paradox where free access defeats the purpose and paid access defeats the regulation.
The revenue is one layer up, in the compliance and attestation work that provenance obligations create — and the single most valuable application, catching distillation, is one the labs will use themselves and never invoice for. When Anthropic ships the detection API, the number that matters is not the accuracy figure. It is the price, because that is the number that reveals which business they think they are in.
Related on explainx.ai
- Anthropic is watermarking Claude text — the announcement and the detection API commitment
- Will every AI model watermark its output? — why open weights cap the whole market
- The case FOR AI watermarks — including why distillation defence funds detection
- How AI text watermarking actually works — why detector access is a security decision
- How to detect a Claude watermark — what you can verify today, for free
- Spotify labels AI artists — a platform buyer for exactly this signal
- LinkedIn's C2PA Content Credentials — the free standard that anchors the price
- EU AI Act and US policy: complete guide — the obligation that creates the compliance market
- LLM text detection with classical ML — the previous detection market and how it went
- What is AI slop? — the demand driver behind platform-scale screening
Primary sources: Anthropic engineering commentary on X (August 11, 2026) confirming a forthcoming text detection API · Anthropic Help Center, "How Claude marks AI-generated content" · EU AI Act Article 50(2) Code of Practice
Accurate as of August 12, 2026. Anthropic had not published pricing, access terms, or a ship date for its text detection API at the time of writing; revenue-pool analysis here is explainx.ai's assessment, not vendor guidance. Follow @explainx_ai for updates.