My token bill runs into the thousands a month, almost all of it from building at work. In the last piece I laid out the ladder, the theory of what the rungs mean. This one is smaller and more concrete. It’s the day itself.
So here’s what the spend bought yesterday.
One major new feature, and a pile of small ones. Analytics for finance to help manage our Claude spend, which means I spent tokens to account for tokens. A strategy deck on how we enable JLL’s global workforce to build this way. A decision memo. An automated maintenance routine for my agents, and an hour improving the loops themselves. And a stretch testing new skills and plugins as they shipped, to see what’s worth keeping.
Several of those were running at the same time, and I was watching none of them closely. That’s the part that surprises people. They picture a coding assistant, one window, one task. A real day is a dozen of them at once, each on its own track.
I stopped hand-writing code years ago. What I do now is set the goal, read what comes back, and point each track at the next thing. The builds are the obvious line item. A feature on one track, a refactor on another, a throwaway prototype to find out if an idea is worth keeping. I move between them.
A lot of the rest is analysis. When I need to understand something, I point a frontier model at the question and let it direct a fleet of cheaper ones. The research for the last post ran that way: it read the engineering corpus, swept X and Reddit, pulled the rankings live, and came back with a few hundred cited findings. A sweep that used to be a week of analyst time now takes one sitting and a few dollars. I run versions of it constantly, mostly on work data.
And a real share goes to explaining things. Leadership briefings, where the job is turning a pile of context into something a room can decide on. Teaching material, to bring other people up the same curve. Talks. Diagrams that make a strategy legible enough to argue about. None of that is the work people imagine when they hear “token bill,” and for me it’s a steady slice of it.
Add it up and it’s all work I was already doing, now running at once. Two things run up the bill. I work in parallel, and I hand the model long, high-level tasks and let it figure out how to do them. I set direction. It does the figuring.
So I read the bill the way the last piece argued you should: as a stack of small machines, each with a job and a cost. The only question that matters is what each one produced. Most months the answer is work I couldn’t have reached otherwise. Some of it I’d have had to hire for, and the work would have waited on the hire. Instead it started yesterday morning.