Most coverage of Anthropic’s S-1 prospectus has gone to the revenue line and the loss line. Buyers should skip both. The pages that matter to a Canadian company paying for Claude are in the Anthropic S-1 risk factors, where Anthropic tells investors, in its own words, that the price, the limits and the models you rely on are all subject to change.
Anthropic filed its registration statement with the US Securities and Exchange Commission on October 26, 2026. Here is the buyer’s reading.
What does the S-1 say about pricing?
A heading in the risk factors reads: “Our pricing structures for our models, platform, and applications change from time to time.” The text goes on: “We have in the past changed, and we expect that in the future we may change, our pricing models.”
It lists the alternatives it may be pushed toward, including “shorter contract durations” and “outcome or task-based pricing.”
The filing is just as direct about access: “We have also changed, and may in the future change, usage limits, rate limits, or model and product capabilities.” And on older models, it warns that customers who built on them “may resist migration to newer versions, and any accelerated or forced deprecation could result in customer attrition, contractual disputes, or reputational harm.”
None of this is a threat. Risk factors are written to protect the company from being sued later. But a vendor that tells regulators it expects to change prices, limits and models is giving its customers fair notice, and most buyers have not written that notice into their contracts. Ask the same question of every vendor’s lenders, because much of the AI buildout runs on private credit.
Show Me the Invoice
Here is where each US$100 of Anthropic revenue went in the first half of 2026, from the S-1’s operating expense table. Revenue for the six months was US$16,286 million, about US$16.3 billion. Operating expenses were US$16,979 million, so the operating loss was US$693 million.
| Measure | Result |
|---|---|
| Where each US$100 of revenue went, six months to June 30, 2026 | |
| Research and development | US$59.81 |
| Sales and marketing | US$23.38 |
| Cost of revenue (compute and infrastructure) | US$14.61 |
| General and administrative | US$6.45 |
| Total operating expenses | US$104.26 |
Serving your requests costs Anthropic under US$15 of every US$100 you pay. Building the next model costs about US$60. On a GAAP basis the company still spends more than it takes in, and even the adjusted margin, which strips out items such as stock-based compensation, is 10%. The S-1 adds that margins “may compress” as new compute comes online ahead of the revenue it supports.
The reading for a buyer: the current rate card is not padded with room for cuts. Price moves from here are as likely to come through new pricing models, limits and tiers as through lower per-token rates. We covered the long-term compute commitments behind this in our Anthropic IPO analysis.
Who has leverage, and who doesn’t?
The S-1 says no single customer was more than 10% of revenue for the nine months to September 30, 2026, but the top five made up about 25%. It also says Anthropic has “agreements with certain large customers which provide flexibility around financial commitments, termination rights, indemnification, and security obligations.”
That sentence is the useful one. Large customers have negotiated terms on exactly the issues the risk factors flag. A mid-sized Canadian company paying by card on standard terms has none of that. At US$3,000 a month, a bill of C$4,211.40 at the Bank of Canada’s October 27 rate of 1.4038, you will not get a bespoke contract. You can still get some of the same protection by choosing how you buy. Read them next to Anthropic’s current commercial terms, which set what a business customer agreed to today.
The bull case
The strongest argument against our reading is the trend. The S-1 shows operating margin moving from minus 184% in the first half of 2025 to minus 4% in the first half of 2026, and revenue up about 14 times. A vendor improving that fast has room to hold prices for customers it wants to keep, and competition from OpenAI, Google and open models limits how far it can push. The risk factors may simply be lawyers covering every base.
We think competition does cap list prices. It does not cap rate limits, model retirements or a move to task-based billing, which are the levers the filing names.
What would change our view
A published commitment from Anthropic to minimum notice periods for price, limit and deprecation changes on standard commercial terms would change our view. So would a quarterly report after listing that shows adjusted margin holding while compute expands.
Run this on your own numbers
Pull your last three months of Claude invoices and list every model name you call. For each one, write down what you would do if it were retired with 60 days’ notice, or if your rate limit were halved. If either answer is “we would stop working,” buy through a channel with written notice terms, such as a cloud marketplace agreement you already have, or keep a tested second model ready. The model deprecation clause we drafted for regulated firms works for anyone. More cost breakdowns are in AI in Business.
FAQ
What do the Anthropic S-1 risk factors say about pricing?
They say Anthropic has changed its pricing models before and expects it may again, including shorter contract durations or outcome or task-based pricing, and that it may change usage and rate limits.
How much of Anthropic’s revenue goes to serving customers?
In the first half of 2026, cost of revenue, substantially all compute and infrastructure, was about US$14.61 of every US$100 of revenue, while research and development took about US$59.81.
How can a smaller Claude customer protect itself?
List every model you depend on, plan for a retirement or a halved rate limit, and buy through a channel with written notice terms or keep a tested second model ready.
