Anthropic has filed to list its shares on the stock market, and the paperwork shows how its customers behave. The businesses that were paying it a year earlier are paying about nine times as much now. That number matters less as a fact about Anthropic than as a description of how AI gets bought.
The bill follows the work
Most business software is sold per seat: you pay for 50 people whether they open it or not. Claude is mostly sold by use. About 80% of Anthropic's revenue in the first half of 2026 was consumption-based, and only 20% came from fixed subscriptions. So the bill tracks how much work the AI does, and that work tends to grow once it proves itself.
For example, a distributor with 300 staff starts by letting Claude sort and answer supplier emails, and the bill is $2,000 a month. It works, so the finance team adds invoice matching, the engineers start using Claude Code, and the operations staff get Cowork for reports and file work. At nine times, the bill is $18,000 a month, or $216,000 a year. That is about $16,000 a month more than the pilot, roughly what two junior office staff cost in many countries. If the AI now does the work of six people, that is a bargain. If nobody measured what it replaced, nobody can say, and that is how a good purchase ends up as an argument in the budget meeting.
One purchase turns into three
Anthropic sells in layers: the models that developers call directly, then apps built on them, such as Claude Code for engineers and Cowork for everyone else. More than half of its customers spending $100,000 a year or more now use both the developer platform and one of the apps, up from about 30% at the end of 2025. Cowork is spreading faster than Claude Code did at the same point after its launch.
The practical result is that a decision made by one team becomes a decision for the whole company. An engineer who tries the API on a company card can end up, a year later, as the reason the sales and finance teams all pay the same vendor.
Cloud computing went the same way. Companies began with a few servers, and as more teams used them, cloud bills rose across the business until finance departments built dedicated teams only to watch them. AI is on the same path, only faster: in a year, Snowflake's customers grew their spending 1.6 times, and Anthropic's grew theirs nine times.
The lock-in is in your habits, not your contract
Most of Anthropic's revenue is consumption-based, and much of it is not covered by long-term commitments. That cuts both ways. Customers can leave as quickly as they arrived, which is a risk for Anthropic. For a buyer, it means the stickiness comes from what you build: prompts, workflows, and people who have learned to work a certain way. Those are what make switching hard, long before any contract does.
So the useful habits are small. Give each team its own budget and look at it monthly. For every workflow, write down the human work it replaces and what that work cost, so a rising bill can be judged against something. And keep the workflows written in plain language, so they could run on another model if a better offer came along. A bill that grows nine times is good news if you can see what it bought.