Claude Code Review – Let AI Find the Bugs You Overlook
Anthropic launched Claude Code Review on March 9, 2026, a new feature for Claude Code that deploys a team of AI agents on every pull request to catch bugs before they reach production. The system is now available as a research preview for Team and Enterprise plan users.
The Problem: Code Review Has Become a Bottleneck
Code output per Anthropic engineer grew by 200% over the past year. That kind of growth puts serious pressure on the review process. As a result, developers are stretched thin, and many pull requests get a quick skim rather than a thorough read.
Before Claude Code Review, only 16% of PRs at Anthropic received substantive review comments. That number now stands at 54%. In other words, more than three times as many pull requests now get the depth of attention they need.
This is precisely the problem the new feature aims to solve. Importantly, it does not replace human judgment. Claude Code Review will never approve a PR on its own, because that remains a human decision.
How Claude Code Review Works
When a developer opens a pull request, Claude Code Review dispatches a team of agents to analyze the changes in parallel. The agents scan for bugs, verify findings to filter out false positives, and rank issues by severity. They then deliver the output directly to the PR as a single high-signal overview comment, plus inline comments for specific bugs.
Furthermore, the system scales its effort to match the size of the change. Large or complex PRs get more agents and a deeper analysis. Small, trivial changes get a lightweight pass. On average, a review takes around 20 minutes to complete.
Real Results and Real Bugs Caught
Anthropic has been running Claude Code Review internally for months, and the data speaks for itself. On large PRs with over 1,000 lines changed, 84% receive findings, with an average of 7.5 issues per review. On small PRs under 50 lines, that drops to 31%, averaging 0.5 issues.
Notably, less than 1% of all findings turn out to be incorrect. That is an exceptionally strong signal-to-noise ratio for an automated system.
One internal case illustrates the value well. A one-line change to a production service looked completely routine, the kind of diff that typically earns a quick approval. Claude Code Review flagged it as critical. The change would have broken authentication for the entire service. Developers had missed this issue entirely, and the engineer confirmed afterwards that they would not have caught it on their own. The team fixed the bug before merge.
Similarly, early access customers have seen comparable catches. On a ZFS encryption refactor in TrueNAS’s open-source middleware, Claude Code Review surfaced a pre-existing bug in adjacent code: a type mismatch that silently wiped the encryption key cache on every sync. It was a latent issue in code the PR happened to touch, exactly the kind of bug that human reviewers scanning a changeset would not normally go looking for.
Cost, Controls, and Availability
Claude Code Review optimizes for depth, not speed. Consequently, it costs more than lighter-weight alternatives like the Claude Code GitHub Action, which Anthropic keeps open source and freely available.
The system bills reviews by token usage, and they generally average between $15 and $25, scaling with PR size and complexity. Admins get several tools to manage spending: monthly organization caps to define total spend across all reviews, repository-level control to enable reviews only on selected repositories, and an analytics dashboard to track PRs reviewed, acceptance rates, and total review costs.
Claude Code Review is now accessible to Team and Enterprise plan users as a research preview in beta. Admins can enable it via Claude Code settings, install the GitHub App, and select the repositories they want to cover. After that, reviews run automatically on every new PR, with no configuration needed from developers.
Source:
claude.com/blog/code-review
