# Cost per successful task: research record

Built: 3 November 2026. Article: `/journal/cost-per-successful-task`.

## Question

What does an AI tool cost per usable result, once the share of attempts that come back right and the minute a person spends checking each one are counted? This is a worked example. There is no outside data; every number is arithmetic on the assumptions below, and the assumptions are printed on each chart.

## Files

- `request.json`: the dataset declaration (no series; rights as declared).
- `build.mjs`: offline rebuild. Holds the assumptions, computes every plotted row, writes `cost-per-task.json`, `validation.json`, the four chart SVGs and `sha256.json`. `node build.mjs` in this directory; no network, no dependencies.
- `cost-per-task.json`: every row of both charts, the assumptions and the formulas.
- `scripts/research/chart.mjs`: the chart renderer. `scripts/research/rasterize.py`: SVG to PNG (2x) and WebP with the site font.
- Tests: `scripts/research/cost-per-task.test.mjs` (in `npm test`).

## Assumptions

| Name | Value | Where it appears |
| --- | --- | --- |
| Example attempt cost | $0.10 | Chart 1 |
| Check | 1 minute of a person's time at $30 an hour, so $0.50 per attempt | Chart 2 |
| Cheaper tool | $0.02 an attempt, 70% of attempts come back right | Chart 2 |
| Dearer tool | $0.20 an attempt, 95% of attempts come back right | Chart 2 |

The prices and rates are chosen to show the arithmetic, not to describe any product. Replace them with your own numbers.

## Formulas

- Cost per successful task: `cost of one attempt / share of attempts that come back right`.
- With checking: `(cost of one attempt + check cost) / share of attempts that come back right`.
- Check cost: `check minutes / 60 * hourly rate`.
- Both formulas assume a failed attempt is caught by the check and tried again, and that attempts fail independently, so the expected number of attempts per successful task is `1 / share that come back right`.

## Headline numbers (from `validation.json`)

- At $0.10 an attempt: $0.105 per successful task at 95% right, $0.125 at 80%, $0.20 at 50%, $0.40 at 25%.
- With the $0.50 check on every attempt: the cheaper tool costs $0.743 per successful task at 70% right; the dearer tool costs $0.737 at 95% right. Both round to $0.74. The price per attempt differs ten-fold; the cost per good result does not.

## Limits

- A wrong answer that is not caught is outside the formula. If the check is skipped, the chart is a floor, not an estimate.
- Setup, training, data export and the cost of leaving the tool are not in the formula.
- Success rates are measured, not read from a price list. A benchmark score is a success rate on someone else's tasks; the rate that matters is the one measured on your own.

## Reproduce

```
cd assets/research/cost-per-task-2026-11-03 && node build.mjs
python3 -I ../../../scripts/research/rasterize.py *.svg
```

`sha256.json` lists the hash of every input and output after a build. Code version: the git commit that last touched this directory.
