American companies are buying more AI than ever and paying less for each unit of it. A token, the small piece of text that AI providers charge by, got 21% cheaper in six weeks, and the number of tokens used hit a record. Total spending has already slipped about 6% from its peak two weeks earlier. The companies selling AI feel it too: payment growth at AI businesses has cooled from a pace of about 140% a year to about 80%.
Part of the falling price is your own choice
Two things pull the average price down. Providers cut their prices, and buyers move routine work to cheaper models, so the price Ramp measures falls even when a company never asks for a discount. The average for OpenAI and Anthropic together is now $0.45 per million tokens. Frontier models cost more than that average and small ones cost less, so what a company pays depends on which model it gives each task.
Take a distributor that gets 20,000 supplier invoices a month and wants each one matched against its purchase order. If every check uses about 5,000 tokens, the job is 100 million tokens a month. At the average price that is $45, and the clerk who now spot-checks one invoice in ten can have every invoice checked. A job like that gets approved or dropped on its cost, and at this price it is hard to drop.
Cheaper units make the seller work harder
Cutting a price by 21% means the seller needs about 27% more volume to earn the same money. In the latest month, volume rose 27% and spending rose 12%, so each dollar bought much more AI than before. For a seller, volume has to keep growing faster than the price falls, and that is a harder race for a small company than for a large one.
The same pattern shows in who is growing. Over a year, the median AI company grew 41%, while the group as a whole grew 103%, which means a few large providers are carrying the average. If a tool you rely on comes from a small vendor, expect it to change its pricing or its product sooner than the big names will. Keep your prompts and your data in forms you can move, which is also how buyers have been getting discounts.
Sellers will also try to sell bigger jobs instead of more tokens. A per-token price that falls every month is a poor thing to build a business on, so expect more AI tools priced per finished task: per invoice matched, per ticket resolved, per candidate screened.
What storage teaches about where this goes
Storage needed twenty years to get about 85% cheaper. The AI price Ramp tracks moved 11% in the last month alone. Expect the habit of keeping everything, and the products that only worked because keeping was cheap, to arrive for AI in a few years instead of twenty.
We reported in November that usage was rising while the bill stayed flat. What is new is that the bill itself has begun to fall, so a budget built on spring prices now buys far more work than it did. The sensible way to plan is to price a finished job, such as one matched invoice, and redo that sum every quarter. The question to keep asking is which work you are still doing by sampling because checking everything once cost too much.