Automated code review just got expensive - or cheap, depending on how you look at it. Anthropic's new Code Review feature for Claude Code sends a fleet of specialized AI agents to analyze every pull request, posting inline comments on the exact lines where it finds problems. It costs $15-25 per review on average.
The feature, currently in research preview for Teams and Enterprise subscribers, works as a managed GitHub App. Install it, pick your repositories, and every new PR gets automatically scanned. No self-hosting, no pipeline configuration beyond the initial setup.
How the Multi-Agent System Works
This isn't a single model reading your diff. Multiple agents run in parallel on Anthropic's infrastructure, each looking for a different class of issue: logic errors, security vulnerabilities, broken edge cases, and regressions that only show up when you understand the surrounding codebase. A separate verification step then checks each candidate finding against actual code behavior to weed out false positives before anything gets posted.
Findings get tagged with three severity levels: bugs that should block merging, minor nits worth fixing but not blocking, and pre-existing issues that were already in the codebase before the PR. Each comment includes collapsible reasoning so you can see exactly why Claude flagged something and how it verified the problem.
The whole process takes about 20 minutes per review. Not instant, but fast enough that results are usually waiting by the time a human reviewer picks up the PR.
$15-25 Per Review Adds Up Fast
The pricing model deserves scrutiny. At $15-25 per review, a team pushing 20 PRs a day is looking at $300-500 daily, or roughly $6,000-10,000 per month. That's comparable to a junior developer's salary in some markets - which is probably the point Anthropic is making.
Choosing the "after every push" trigger mode multiplies that cost by however many commits land on each PR branch. A PR with 8 pushes could run up $120-200 in review costs alone. Anthropic lets admins set monthly spend caps, which feels less like a feature and more like a necessity.
For comparison, GitHub Copilot's code review (included with Enterprise at $39/user/month) and existing static analysis tools like SonarQube cost a fraction of this. Anthropic is betting that deeper, context-aware analysis justifies the premium.
Customization Through Markdown Files
The most practical detail: you control what Claude flags by adding a REVIEW.md file to your repo root. Want it to always check that new API endpoints have integration tests? Add a line. Want it to skip generated code under /gen/? Add another line. The system also reads your existing CLAUDE.md files, so project-level coding standards automatically become review criteria.
This is a smart design choice. Instead of building a complex rules UI, Anthropic made configuration a markdown file that lives in version control alongside your code. Teams can review and iterate on their review rules the same way they review code.
One notable limitation: Code Review is not available for organizations with Zero Data Retention enabled, which means some security-conscious enterprises that might benefit most from automated vulnerability scanning can't use it.
The feature focuses on correctness by default - actual bugs that would break production, not style nitpicks or missing test coverage. That's the right starting point. Most developers' complaint about automated review tools is noise, not silence. Anthropic seems to understand that a tool posting dozens of formatting suggestions on every PR would get disabled within a week.