# Cost per solved task: the one-sheet agent cost check (free sample)

A free sample from Agent Cost Levers by PaperChaseWebb. Works with Claude Code, Codex, Cursor or any agent tool that reports usage.

Most agent cost advice starts with the model price. Start with what you control: how many tasks actually get done for the money.

## The number

**cost per solved task = total spend on a batch / tasks that passed a check you wrote before the run**

- A task is solved only when a check says so: a test, a script, a lint pass, a two-minute reviewer checklist.
- Failed and abandoned runs stay in the total.

## The sheet

| Task | Run | Input tokens | Cached input | Output tokens | Tool calls | Passed (Y/N) | Cost |
|---|---|---|---|---|---|---|---|
| | baseline | | | | | | |

## Three rules

1. Use ten real tasks from your own week, with a frozen starting commit for each.
2. Change one thing at a time and rerun the same ten.
3. Keep a change only if cost per solved task drops and the pass count holds.

## The check prompt

"Here are the baseline and the new run for the same ten tasks [paste both]. Compute cost per solved task and the pass count for each. If the pass count dropped, say so first. Do not call anything a saving if quality fell."

## In the full pack

The five levers that move the number (thinking depth, context curation, direct tool routing, cache and reuse, verified stop), a rules block for CLAUDE.md, AGENTS.md and Cursor, a claim discipline checklist for reporting savings, and 10 prompts.

License: personal and commercial use.