Example
If the all-strong estimate is $0.42 and tokenpatch execution costs $0.08, the estimated savings ratio is 81%.
Cost calculator
Estimate AI coding token cost per applied patch.
The useful unit is not one API request. For AI coding, the useful unit is an accepted patch. tokenpatch reports executor usage, estimated all-strong baseline, actual executor cost, and savings ratio after each bounded run.
| Metric | Formula | Why it matters |
|---|---|---|
| All-strong baseline | Strong model input/output tokens x strong model price | Approximates what the task would cost if the expensive model did all implementation work. |
| Executor actual | Executor input/output tokens x executor price | Shows the lower-cost implementation spend for the bounded patch. |
| Savings ratio | (baseline - actual) / baseline | Turns token savings into a task-level percentage that is easy to compare across runs. |
| Cost per applied patch | Actual executor cost / accepted patch count | Filters out vanity request metrics and focuses on useful code changes. |
What to track
Track baseline, executor usage, accepted patch status, validation result, and whether the diff stayed inside allowed files.
What not to overfit
Do not optimize only for cheapest tokens. A failed patch can cost more than a slightly stronger executor that lands the change.