Anthropic’s 5% to 7% revenue line is a detection problem, not a scandal
The sentence in the S-1 should not be understood to mean that “users who broke the rules contributed 5% to 7% of revenue.”
For the nine months ending September 30, 2026, about 5% to 7% of revenue came from customers who were later found to have broken usage policies or business terms. Using $16.286 billion for the first half of the year and the media-reported Q3 figure of $16.8 billion that was shared with investors, revenue for the nine months comes to about $33.086 billion. Between 5% and 7% of that amount is roughly $1.65 billion to $2.32 billion. The S-1 explains when this revenue was recognized: the company had already provided the service before restricting or closing the accounts. The money came in during the time between the rule-breaking, its detection, the investigation, and the account's closure.
For the same demand, if the problem is detected on the first day, this revenue is close to zero. If it is detected only after three months, the company collects three months of payments. So the 5% to 7% figure is mainly a financial estimate based on how long detection and account closure take. “Revenue earned while violations go undetected” is more accurate than “revenue from users who break the rules.”
This estimate is not straightforward. It needs to be separated into at least four different sets of accounts.
First set of accounts: accounting and disclosure are different
The statement that “the service was provided and revenue was recognized under the rules” is not the main point of debate. The concern is the wording about changing the estimate when detection methods improve. The S-1 says the estimate may change because better monitoring will uncover usage problems that were not found before. This wording leaves room for legal adjustments, but it does not mean that revenue already recognized will be reduced afterward.
The first change is more likely to be a change in how the company describes its revenue. Some past revenue may be relabeled as “revenue later found to involve violations,” making the share appear larger when viewed afterward. What actually changes the income statement includes refunds, credits or compensation, refunds of unused prepaid balances, and unpaid bills from customers who pay after using the service. The S-1 does not promise to revise past revenue, and it does not give figures for refunds. Interpreting “the estimate may be adjusted” to mean “past revenue will be recalculated downward” goes beyond what the prospectus says.
Changing the categories used to describe revenue and changing the amount of revenue recorded in the accounts are two separate risks. Mixing them makes the analysis less legally precise.
Second set of accounts: the ratio combines six factors
The 5% to 7% figure combines several things: how much paying business comes from customers who break the rules; the type of violation, including content violations, identity or regional restrictions, and model distillation; how customers pay, whether through usage charges, subscriptions, or advance payments; the actual time between discovering a problem and closing the case; whether unused balances are refunded when an account closes; unpaid bills from customers who pay afterward; and new revenue from customers who register again.
The basic idea is that earlier detection brings this revenue closer to zero. But a high ratio does not directly prove that the company's system for controlling risk is ineffective. A better way to understand it is that the ratio reflects the mix of high-risk customers, the time it takes to deal with violations, and the way customers pay. If API resellers, bulk accounts, and demand related to model distillation make up a large share of the customer base, that business is naturally more short-lived, relies more on advance payments, and makes it easy to register again. As a result, more revenue can build up before violations are detected. This does not excuse weak risk controls. It means the ratio alone cannot explain very much.
One accounting detail is important. If unused prepaid balances are refunded when an account closes, the revenue the company actually keeps will be lower than the revenue it has already recorded. If those balances are not refunded, the amount kept will be closer to the full amount earned during the detection delay. The S-1 does not explain how these cases are divided. The 5% to 7% figure is therefore closer to an upper estimate than an exact measure of revenue kept from violations.
Third set of accounts: closures and new registrations do not automatically create extra net revenue
The figures in the transparency report stand out. In the first half of 2026, 11.4 million accounts were banned, 398,000 appeals were received, and 42,000 bans were overturned. That means about 10.5% of appeals succeeded. The figure of 11.4 million bans cannot mean that 11.4 million different people or businesses were involved. Resellers manage large numbers of accounts and switch to new ones when accounts are banned. This is a known way of operating in the industry.
The claim that the same demand is repeatedly charged needs three important qualifications.
First, the number of accounts is not the same as the number of paying customers. Neither the S-1 nor the transparency report says how many of the 11.4 million accounts were free accounts, trial accounts, or fake accounts registered in bulk. Many bans may generate no revenue at all and may simply be actions taken to control risk.
Second, revenue from registering again is not pure extra income. The platform repeatedly collects subscription fees or advance payments, but it also faces the costs of controlling risk, payment chargebacks and refunds, payment processing, and customer service for appeals. If the same demand is split across ten short-lived accounts, the amount of revenue recorded may rise, but the amount the company keeps from that demand may not. Treating repeated charges as direct profit exaggerates the economic value of registering again.
Third, the 10.5% success rate for appeals shows that some accounts are banned by mistake. But this does not prove that “many banned accounts belong to normal paying customers.” People who take the time to appeal and provide proof of their work are likely to be frequent, heavy users. The 398,000 appeals represent only a small part of the 11.4 million bans. The available data cannot show how many of the accounts that did not appeal involved genuine violations, failed identity checks, or mistaken bans that users simply gave up challenging.
Registering again splits the same demand across several short-lived accounts and can lead to repeated revenue being recorded while violations remain undetected. To work out the net benefit, the company must subtract refunds, chargebacks, and risk-control costs. Account closures and new registrations cannot be treated as a simple way to increase revenue.
Fourth set of accounts: some regional restrictions are really about identity and ownership
In September 2025, Anthropic updated its terms. Organizations that are more than 50% directly or indirectly owned by Chinese capital fall under the restriction, whether they are registered in mainland China, Singapore, or the United States. The S-1 does not break down the 5% to 7% figure by type of violation. It is reasonable to think that a substantial part does not come from users creating illegal or abusive content. Instead, the issue may be their identity or ownership structure, which places them in a region where the service is not supported.
This type of demand has different features. Writing code, editing documents, and doing research are ordinary work tasks. People doing them may be willing to pay a lot and have few alternatives. This is a key reason the reseller business has continued to exist. For these users, a ban is not punishment for breaking a content rule. They lose access because their identity or ownership does not meet the regional requirements. They have a strong reason to switch to another account and keep paying. This increases the share of revenue earned during the detection delay and weakens the idea that bans reduce revenue even when demand remains strong.
This point must be clearly described as a reasonable inference, not a fact disclosed in the S-1. Saying that “much of the 5% to 7% comes from regional restrictions” goes beyond the evidence in the documents. It is more defensible to say that “a considerable share may come from people who do not meet identity or ownership requirements, because actual demand and willingness to pay make it hard to assume that a blanket ban will work.”
Disclosure is a legal act; materiality does not depend on a strict 5% cutoff
A source of revenue that makes up 5% to 7% of the total, that the company is actively trying to reduce, and that involves account bans and regional restrictions is the kind of risk that would normally need to be disclosed under common standards for judging what is important to investors. Under U.S. securities law, there is no simple rule that a risk becomes important only when it reaches 5%. The judgment depends on both how likely the risk is and how large its effects could be. This ratio, together with the forward-looking statement that detection and account closure measures were strengthened in August and September, that the share is expected to fall, and that this may affect total revenue and growth, provides a fairly complete warning about the risk in advance.
Its first purpose is to help defend the company legally. If growth in Q4 or later periods looks weak because the company is reducing revenue linked to violations, the prospectus has already warned that this may lower future revenue and growth. If lawyers representing investors in a class-action lawsuit examine the document line by line, this warning gives the company a ready-made basis for its defense. The second purpose is to support its public relations and safety message: “We banned so many accounts that it hurt revenue, but we banned them anyway.” Regulators and business customers may read this as a statement about the company's approach. For investors, it is mainly a risk that could slow growth and has already been disclosed as such.
Disclosure is first a legal act and only second a public relations act. Reading it as a declaration of conscience misses the main purpose of the document.
The October 12 announcement: users begin to bear some of the cost of mistaken decisions
The October 12 announcement is consistent with the statement in the S-1. In specific areas, such as model distillation, Anthropic resumed charging for requests that its safety system blocked in advance, starting September 24. The company said it was doing this because the rate of false positives was below 0.1%. It also acknowledged that users would be charged for normal requests that were wrongly blocked.
The rule's basic effect is that people who break the rules bear the cost of their actions, while users also bear part of the cost when the system makes a mistake. For the platform, this means taking the cost of risk control into account within its own pricing system. If you send a request that may trigger a block, you may have to pay for the blocking action even if the request is later judged to be harmless. The platform no longer bears that cost by itself.
Three qualifications are important. First, 0.1% is the rate of false positives per request, not per account. When billions of requests are involved, even that small percentage can mean a large number of mistakes. Second, the charges apply only to specific areas, such as model distillation. The company has not generally resumed charging for all blocked requests. Third, the announcement does not explain whether the process for appealing a mistaken block and getting a refund is equally fair to users. This rule may add to revenue in the short term, but its effects on fairness and customer relationships cannot be judged by revenue alone.
Sensitivity analysis: fill in the missing base before judging Q4 growth
The main conclusion still holds. The 5% to 7% figure represents revenue earned while violations go undetected, not fines charged to a group of rule-breaking users. Better detection will reduce future growth. Disclosure is first a legal act. But three points need to be treated carefully.
Do not treat 5% to 7% as a clear measure of how well the company controls risk. It combines the types of customers, payment methods, the time needed to handle cases, and refunds or unpaid bills.
Do not treat repeated charges from new registrations as a net benefit. The number of accounts is not the same as the number of paying customers, and repeated revenue must be reduced by refunds, chargebacks, and risk-control costs.
Do not treat the share of revenue linked to regional restrictions as a fact disclosed by the company. It is reasonable to infer that people who do not meet identity or ownership requirements account for a considerable share, but the amount cannot be measured from the available information. It must not be presented as something the S-1 says when it does not.
Here is what the sensitivity analysis shows in simple terms. If the detection period falls from three months to one month, the revenue kept from the same demand during the delay would fall to roughly one-third of its previous level. The reduction would be about two-thirds of the original delayed revenue. If detection improves from three months to almost zero, the revenue kept would approach zero, and the reduction would approach the full original amount. If detection was strengthened only in August and September, the figure for the first nine months would fall only partly. The size of the change would depend on when revenue is recognized in Q4 and how much is refunded.
Using a total revenue base of $33.086 billion and a range of 5% to 7%, revenue earned during the detection delay is about $1.65 billion to $2.32 billion. If the reduction in Q4 is $0.5 billion to $0.8 billion, the effect on Q4 year-over-year growth is roughly the amount of the reduction divided by revenue in Q4 of the previous year. If that previous-year Q4 revenue was between $15 billion and $17 billion, the reduction would lower growth by about 3 to 5 percentage points. If the base was $20 billion, the reduction would be about 2.5 to 4 percentage points. The actual base has not been disclosed, so it remains unknown. Before deciding how to interpret Q4 growth, fill in the previous year's Q4 revenue base and the refund rate.