On Monday, Anthropic cancelled a price increase. I want to type that sentence again, slowly, because nothing like it has ever happened to me in the physical economy. A supplier scheduled a 50 percent increase, published the effective date, and then called it off.
The details. Sonnet 5 launched in June at $2 per million input tokens and $10 per million output, labeled introductory pricing through August 31. The jump to $3/$15 on September 1 was confirmed in writing recently enough that I built a budget around it. This week the introductory rate became the permanent rate. The increase is off. The pricing page now just says what it says, with no asterisk counting down to Labor Day.
I had a line item for this. An actual row in an actual spreadsheet. I meter tokens per workflow because the workflows bill real customers: estimate generation for the estimating tenants, call summaries coming off the dialer, the nightly anomaly pass over payroll before anything gets my signature. A generated estimate costs me about four cents of tokens today. The September rate would have made it six. Four cents to six sounds like nothing until you multiply it by every estimate every tenant runs, at which point it becomes a real number with its own row and its own opinions.
In my other life this is called a change order, and I have written about those. What I have never seen is a change order in the customer's favor that the customer didn't fight for. No lumber yard has ever called me to say the increase is off and the old price stands. Prices ratchet. Ratcheting is the entire personality of a price.
Introductory pricing is a gym membership. It exists to let you build habits at one number and get billed at another. Everyone who wired Sonnet 5 into production in June did the June math knowing the September math was coming, the way you join a gym in January knowing what happens to the rate when the resolution wears off. Making the intro rate permanent is the gym deciding to stay cheap. Gyms do not do this. Something changed.
My read on what changed: models are shipping like software patches now, several a month, from more vendors than I can keep in my head. When the capability gap between releases narrows, the number on the pricing page becomes the benchmark that matters most. And businesses were visibly budgeting for September 1, which is poison for the thing that actually decides these markets. I keep saying adoption is the deliverable. Predictable unit cost is the schedule of values underneath it. Nobody standardizes on a meter that announces it will spin faster in the fall.
What I actually did about it: almost nothing, which is the good part. The margin math from June still holds, and boring is the best available news in a cost model. The routing stays as it was — the cheap model summarizes calls, the good model touches anything a customer signs. I wrote yesterday that agents are done piloting, and production systems get to be picky about their suppliers. Because here is the fine print I carry around: permanent, in an API doc, means until the next doc commit. A price that can be lowered by a web page update can be raised by one.
No supplier in the physical world has ever un-raised a price on me. The first one to do it sells thinking by the pound.
Still. I have negotiated steel, lumber, diesel surcharges, and dumpster fees, and I got this concession by doing nothing at all except being one of many customers with a spreadsheet. Competition did the negotiating. It is a strange feeling, sitting on the side of the table the market fights for.
The September 1 row stays in the spreadsheet. It reads zero now. Every budget deserves one monument.