Ask every time, or build once?

The same question asked to an agent every working day, compared with having the agent write a script once and running that script for free. The presets are measured runs on Claude Sonnet 5.5, not guesses. Change anything.

Scenario

Asking the agent, per run

Model price, $ per million tokens

How often

Building it once

People's time

Asking, per year
Building, first year
Pays for itself
Runs a year
the script part answers the same way each time

Cumulative cost over a working year

Ask the agent every time Build once, run the script

Measurements, scripts and raw usage logs: the ask-once-build-once repository that goes with the post. Prices are list prices on 15 October 2026. Tokens counted by the Claude Code CLI (claude -p --output-format json).