What the Monthly Bill Made Us Think
When we counted what the Dream Team bots do in a day, most of it was simple repetition: tidying titles, detecting languages, checking formats. Yet every single call was going to the largest model. It was like handling a new hire's errand in the executive boardroom.
The Experiment: Three Tiers
For one week, we sorted calls like this.
Light work: small model (classification, format checks, summary drafts)
Regular work: mid-size model (translation drafts, caption writing)
Hard work: large model (strategic judgment, final review)
There was only one rule: 「when unsure, go one tier up.」
Results
Out of every ten calls, seven landed in the light tier, two in the regular tier, and only one reached the large model. The bill felt like it dropped to under half. What about quality? The small model did miss a few format issues, but the final large-model review caught every one of them.
Strategy Notes for Small Teams
• AI running costs depend on 「who you assign the work to」, not just the price per model.
• In the early startup stage, split and label your tasks before you pick a model.
• Keeping only the final review on the large model preserves your quality insurance while lowering the bill.
It is not about being stingy with expensive tools. It is about recognizing the moments that truly need them.