Cost Per PR Estimator

How to read the results

Cost per PR is half of an ROI equation — the other half is what a PR would have cost in engineering time without the tooling, at whatever loaded hourly rate you use internally. Count merged PRs in the denominator, not opened ones, and treat the output as an estimate to verify against your provider invoice.

FAQ

How is the estimate computed?

Straight arithmetic on your inputs: session tokens × sessions per PR × per-Mtok rate / 1,000,000. Worked example: at $3/M input, a PR that consumes 1M input tokens across its sessions costs $3.00 on the input lane, plus output at its own rate.

What are the limitations?

Attribution is fuzzy: sessions span multiple PRs, exploration and abandoned branches consume tokens that never merge, and review or debugging sessions blur the boundary. The number is directional — useful for trends and comparisons, not cost-accounting.

What's a "good" cost per PR?

There's no universal benchmark, and we won't invent one. The comparison that matters is internal: your cost per PR against the loaded cost of the engineering time it replaced, tracked over time as your setup changes.

How do I lower cost per PR?

Reduce the tokens each session burns before real work starts: standing context like CLAUDE.md and MCP tool definitions ships with every request. The 100-config analysis found a median config of 964 tokens / 92 lines — and that waste concentrates, so a few fixes go a long way.

When should I use this vs the full audit?

Use this to report a unit-cost number. The full stack audit finds the setup-level waste that inflates it — the audit output feeds directly into a lower tokens-per-session input here.