Review Bill: predict what an agent code review will cost before the PR lands

I found no tool that predicts what an AI code review of a specific PR will cost before it runs, priced for the reviewer you actually use (Claude Code Review, CodeRabbit overage per file, Greptile credits, Copilot credits). What exists falls into three groups. Vendors show averages and spend caps after the fact. Calculators and blog tables estimate at the team level. CI cost gates like ci-guardrail, tokentoll, TokenFlow and TokenDam measure the LLM spend of the app or of the agent that wrote the code, not of the reviewer. The code-health score on its own is crowded. pr-health-agent, Typo and react-doctor already do it, and established code-quality vendors give it away, so it shouldn't be what you sell. Your opening is the cost forecast for each PR. Make it accurate per vendor and use it to drive action: hold back or downgrade the review on PRs likely to be expensive, warn when a PR is likely to land in an outlier band like the reported $107 reviews, and suggest splitting the PR to save money. Calibrating on public agent PRs is a reasonable start, but an edge that lasts would come from adjusting the model against each team's real review bills. The risks are that the cost-gate tools and vendors like repowise and Codacy, who already write about cost per PR, are a short step away. Anthropic or GitHub could also add a pre-review estimate in their own products. And if review pricing moves to flat per-seat plans, the problem disappears, so this is a narrow, early opening rather than open space.

• Claude Code Review (Anthropic)

  This is the service whose cost you'd be predicting. Anthropic says reviews average $15 to $25 and scale with PR size and complexity. Admins get a monthly spend cap and a choice of when reviews trigger, but I found no estimate shown before a review runs. One user reported paying $214 for two reviews, which shows how far bills can stray from the average.

• GitHub Copilot code review billing

  A third-party guide says GitHub gives rough per-review credit ranges by tier: $0.05 to $1 on Lite and $0.25 to $5 on Balanced. These are published ranges, not a forecast for each PR.

• Arnica AI Code Scanning Cost Calculator

  A free web calculator that estimates the yearly bill from team size and PRs per month, using fixed per-PR figures. It works at the team level, not per PR, and doesn't plug into CI.

• ai-cost-estimator.com

  Blog content with tables of token use and cost by PR size. It's the same idea as your calibration, but as articles rather than a tool that comments on PRs.

• repowise

  A vendor writing about review cost per PR at agent volume, plus a 200-PR budget for Claude Code Review. They clearly care about this topic and could add a feature for it.

• ci-guardrail (wartzarbee)

  A GitHub Action that estimates how much a PR changes token cost, comments on the files responsible, and can fail the build or only warn. It's aimed at the runtime token cost of agent code, not the cost of reviewing the PR, but the mechanics are almost the same.

• tokentoll

  Scans Python and JS/TS for LLM API calls and gives each PR a PASS, WARN or FAIL against a policy file. It covers app LLM spend, not review spend.

• TokenFlow

  Reads local logs from Claude Code, Codex and Cursor and posts the agent's cost on the PR as a comment and status check, with budgets. It reports what was spent writing the code, not what reviewing it will cost.

• TokenDam

  A linter for LLM token spend, with a GitHub App that fails builds when costs go up. It launched recently and has little traction.

• pr-health-agent

  A Python command-line tool that scores PR health out of 100 across 13 categories, including complexity, and posts the result as a PR comment. This overlaps with the code-health score half of your idea.

• Typo PR Health

  A commercial PR health score weighted on diff size, files changed and cyclomatic complexity. It also covers the code-health half.

• react-doctor GitHub Action

  A health-score Action (0 to 100) for React codebases that comments on PRs. It only covers React.

Checked 2026-10-16