Why your AI agent bill keeps climbing and how to cut it

The first month of the Telegram bot was a lesson in model bills. Social Command Center turns one idea into four platforms of copy plus an AI image, and the bill climbed until I noticed it on a coffee refill — because I was pointing every task at the strongest model I knew.

One question fixed it, repeated for every step of the pipeline: what does this step actually need? Writing LinkedIn copy does not need the same model that draws text on glass.

Model splitting, done on purpose

I split the pipeline into two paths. Copy goes through a cheap text endpoint — the kind that costs fractions of a cent. The image model fires only when a visual is part of the job, and even then I pay for the strongest tier only when the art strays from the brand guide.

Result: a four-platform publish runs about nine cents when it needs text baked into the image, and one cent when it does not. Total running cost, server and all, sits around $15–20 a month. MoneyGuard keeps the same discipline on the other side — a beta running on DeepSeek instead of a flagship model, holding the entire bill in single digits.

Do not pay for cloud you can run

The expensive, boring habit is paying for infrastructure because it's convenient rather than because you need it. The bot lives on a single droplet under systemd, with SQLite as the store and Bun as the runtime. On Melow, the vector database is bound to localhost — Qdrant never leaves the machine, which removes the managed vector cloud fee and a whole security-review question in one move.

Measure actions, not API calls

The metric that actually changed my decisions was cost per successful action — "post published", "record updated" — not cost per API call. Counting calls hides the waste; counting deliverables exposes it.

The unexciting truth: the expensive endpoint stays the default because integration code was written once and never revisited. The fix took an hour — a cost-aware model selector and a log line that prints what each publish path cost. Now that log sits in the terminal and reminds me what every action is worth paying for.

If your agent has a monthly bill that makes you wince — read how I cut mine — run the same audit: cheap models for the steps that are cheap, expensive models only when the output is on display, and local storage anywhere the cloud is just convenience.

Building something like this? Tell me what you need and I'll scope it.