Even Anthropic's customers are routing by price, not power

Enterprise AI spending has quietly split from the capability race. Inside the company most associated with pushing that race forward, customers vote with their budgets, and the votes are going to the older, cheaper model.

Simon Willison's link blog, citing FT reporting on Ramp's billing data, captures the gap with unusual clarity. Fable 5, Anthropic's newest and most expensive offering, captured about 8 percent of Anthropic spend in its first partial month on the market, against Opus 4.8's 28 percent. Three of the four premium-priced models in the lineup sit in single digits. The company is still growing fast: annualized revenue jumped from roughly $47 billion in May to about $65 billion in July, with about 6,000 customers spending $100,000 or more a year.

The naive read is that Fable 5 is failing. The cleaner read is that "best" and "bought" are no longer the same word. Capability leadership, in other words, has decoupled from customer adoption, and the vendor's own revenue mix is the most legible proof. OpenAI's July quarter, where GPT 5.6 jolted performance after a sluggish start, suggests the same pattern travels across labs.

The mechanism travels further than this case. When a flagship model loses share to its cheaper predecessor inside the same vendor's stack, the story is not about model quality. It is about who carries the inference bill and how they route it. Cost-led adoption is now the default setting for enterprise AI, and Anthropic's own billing data is the receipt.