Somebody on your team asked for "the best AI model." Nobody asked what it's b...

Somebody on your team asked for "the best AI model." Nobody asked what it's best at. That unasked question is most of your bill.

To be fair, the premium model earns its keep in one spot: the long, genuinely hard job with a deadline (the kind where the cheaper models get there too, just a few months later than you needed them). DoorDash runs it exactly this way: cheap model for the routine work, frontier model for the tasks that would otherwise eat an expert's whole afternoon. So this isn't "always buy cheap." It's that almost nothing you do all day is that job. You're paying premium prices across the board anyway.

This same mistake has a twin, and it's living in your codebase. The AI coding tools your team picked up are shipping features noticeably faster (everyone can see that part). What nobody put on the slide is the second bill that rides along with the speed.

Researchers looked at 441 codebases. The only thing separating the teams who doubled their mess from the teams who didn't was a few pages of committed instructions (the house rules, handed to the AI up front). Written once, and three out of four teams never touched them again. An afternoon of setup, skipped, then billed to you every month after.

Same shape, twice. Two different bills, one cause: a default nobody went back and questioned.

The fix for both is the same: measure it on your own work. Take a job you actually do every week, run it on the cheap model and the expensive one, and see if you can tell the difference. Most of the time you can't (and you just cut that task to a fifth of the cost). Same with the code: write the house rules, then watch whether the mess actually shrinks. Your tasks, your numbers.

The AI was never the expensive part. Not looking was.