Anthropic wants Claude to do more than just write code. The company has introduced Code Review, a new feature aimed at catching coding mistakes before software goes live, expanding Claude’s role from generator to reviewer.
According to Anthropic, the launch brings a more hands-on layer of AI oversight into the coding process as developers race to ship more software faster. The feature is rolling out in research preview through Claude Code.
Code Review splits the work, then compares notes
Code Review does not rely on a single AI pass.
The feature sends out a team of agents the moment a code change is opened for review, with each checking for bugs simultaneously. Those findings are then cross-checked to filter out false positives before the system sorts them by severity so the most serious risks rise to the top.
What comes back is meant to be usable, not overwhelming. Instead of dumping a messy stack of warnings on developers, Code Review returns a single high-signal summary comment, along with inline notes attached to specific parts of the code.
The review also varies with the size of the update. More complex code changes get more agents and a deeper inspection, while minor edits get a lighter pass. The average review takes around 20 minutes.
The internal numbers are already piling up
According to Anthropic, the feature is already used on nearly every internal code submission, a sign that this is no small in-house trial. The company says the share of code changes receiving substantive review comments climbed from 16% to 54%.
The results were strongest on larger code updates, with Anthropic noting that those reviews were far more likely to surface problems than smaller edits. That is where the tool may matter most, since large changes are harder for people to check thoroughly and more likely to conceal expensive mistakes.
Two bug catches show the real value
Anthropic’s clearest example began with a one-line change that barely looked risky. It was the kind of routine edit that could easily have picked up a quick approval and moved on. Instead, Code Review flagged it as critical because it would have broken authentication before the change was merged.
The second example came from early access user TrueNAS and points to a different kind of risk. During work on ZFS encryption, Code Review surfaced an older bug nearby: a type mismatch, meaning two pieces of code were handling data in incompatible ways, which was wiping the encryption key cache on every sync.
Anthropic’s point is that some of the most serious problems are easy to miss and do not always appear in the exact line a developer set out to change.
Who gets it, and what it costs
Code Review is not the budget option. It is built for deeper inspection, with token-based pricing that typically runs $15 to $25, depending on the size and complexity of the code change. Anthropic presents it as the heavier-duty counterpart to its lighter Claude Code GitHub Action.
There is also a practical layer for teams managing rollout and spend. Admins can set monthly organization caps, switch reviews on only for selected repositories, and use an analytics dashboard to track how many reviews ran, how often findings were accepted, and what the reviews cost.
Access is currently limited to Team and Enterprise plans, with setup requiring admins to enable the feature in Claude Code, install the GitHub App, and choose which repositories to cover.
OpenAI is expanding Codex into security with a preview designed to help teams validate and remediate issues faster.