AI cost per pull request

The question

What does AI-assisted software development actually cost?

Hypothesis

Token/API spend plus review and rework time can be measured per PR if model calls, durations, and outcomes are logged consistently.

Method

Instrument representative PRs, record model, tokens, duration, human review time, and rework, then calculate fully loaded cost per merged PR.

Early signal

Task-level costs are now measured in production (Ondari): token counts priced server-side per model, e.g. a 2,400-input / 800-output claude-sonnet-4.5 task at $0.0192, reproducible to the token. The missing piece is attribution — grouping those measured tasks under the PR that caused them. PR-level instrumentation is the next step.

Outcome

Partial: unit economics proven at task level; PR aggregation in progress. This entry is ACTIVE.