The router is not choosing the model, it is choosing what the model sees

Stripe agreed to buy OpenRouter for over seven billion dollars, per the Bloomberg report the newsletters were chewing on this week, and the follow on commentary went straight to the margin question. A layer taking five percent of spend is a wonderful business right up until somebody offers the same thing at zero.

Routing gets sold as arbitrage. Same task, cheaper model, keep the difference.

The more interesting version showed up in an interview with Glean’s Arvind Jain a day later. He says the reason enterprises want routing is cost, plainly, and that per user spend has climbed because the strongest models charge more per token and people point them at much longer tasks. Fine. But then he describes what their system does before any of that. A model called Waldo decides how to break the question down, which tools to use, what to read next, and when there is enough evidence to hand off to an expensive model. His line is the one worth keeping: a cheaper model with better context can beat a frontier model loaded with irrelevant data.

The vault’s own routing notes arrive at the same place from the other end. Cheap or local models for the iterative execution loops, frontier tokens preserved for planning and verification, and never hand project ownership to a mini model.

So the decision that matters is not which model gets the request. It is how much work happened before any model got the request. Assembling the raw materials is not a routing feature. It is most of the job, and the router is just where the bill shows up.