stripe bought a router for $7.5b. i ran one for a year, and routing was never the hard part.

the big number this week is $7.5b for a company whose product picks which model answers your request. smart purchase, and here is why: stripe did not buy intelligence, they bought the meter. 10 trillion tokens a day flowing through one pipe is a phenomenal billing and data asset. but after running a multi-model router daily for about a year, here is the part i cannot unsee: routing is a pricing problem wearing an intelligence costume. picking the cheapest model that can probably handle the task is arithmetic with extra steps. the actual hard part is the thing nobody bills for: verifying what came back. from my logs, roughly 3 in 10 confident outputs do not survive cross-examination (a second model that only reads failures, plus a human spot check). routing saves money on the easy 7. what does the meter do about the wrong 3? nothing today, because correctness is not metered. the next billion-dollar layer is the tribunal: proof, provenance, and a record of what got thrown away. open question for anyone running production traffic: if your router quietly downgrades to a smaller model on tasks that look easy, how would you know before your users do? (the long version of the year of logs: https://www.reddit.com/r/artificial/comments/1x9nei2/)