The Biggest Model, Once a Day
At first, every bot used the smartest model available. It felt like the safe choice. Then I opened the usage table at the end of the month and found that simple tasks, the ones that run dozens of times a day, made up most of the cost. We were using the most expensive model to sort, shorten, and format text.
Three rules, that was all
1. Sorting, summarizing, and formatting go to the small model
2. First drafts go to the mid-size model
3. Only strategic judgment and final review go to the large model, and only once a day
Each bot now writes 「which tier is this job?」 on the first line of every task. If no tier is written, the mid-size model is the default.
Three weeks of measurements
• Token usage down 41%
• Revision requests went from 6 a week to 5, practically unchanged
• Responses in the small-model tier got noticeably faster
An unexpected finding
The bigger win was not the cost. Writing the tier every time made the difficulty of each job visible, and less than 10% of our work truly needed the large model. The rest was being handled expensively out of habit.
Notes for small teams
• The first step in cutting fixed costs is not dropping tools, it is tiering how you use them.
• Splitting work between a Windows device and a Mac mini also spreads out when the large model gets called.
• Before agonizing over a revenue model, learn the unit cost of handling one task. With that number, you can set pricing and decide how fast to scale.
One line for today: with smart models, spending them wisely beats spending them often.