The expensive model was five times the cost for a job that never thinks twice

Every job I hand to AI could run on the most capable model available. Almost none of them need to. The real cost of running AI on autopilot hides in that gap, between a model that has to reason through something new and one that is just following the same steps it followed yesterday, and I found it sitting inside the system I run the most often: the one that writes captions for the short videos I post almost every day.

That job used to run on Claude’s biggest model, Opus, the same one I reach for when something genuinely needs to be thought through. Captions are not that kind of work. The job reads a transcript, picks an opening line, and moves to the next video. It does the same shape of thinking five times a day, every day.

Running the biggest model on a job like that is like hiring your most expensive consultant to file paperwork. The work gets done. It also gets billed like it needed someone who cannot be replaced.

Claude’s smaller model, Sonnet, costs $2 per million tokens in and $10 per million out, a token being the small chunk of text AI companies bill by, the same price it has always been. Opus costs several times more per token on top of that, and for a job that repeats the same steps every run, that gap compounds every single time it fires.

I moved the captions job to Sonnet. Nothing about the output changed. Opus had been costing about five times as much per run to produce a caption a cheaper model wrote just as well.

The judgment call was never “always use the cheap model.” It is knowing which jobs actually need the expensive one. A job that has to notice something nobody told it to look for, or make a real call, earns the bigger model. A job that runs the same checklist every time does not, no matter how natural it feels to default to the best tool sitting in front of you.

If anything you run on AI repeats the same shape of work over and over, ask what it is actually being asked to do each time it fires. Repetition and judgment are not the same job, and paying for judgment on a repetition task is a bill that never needed to exist.