Code Review is generally available: what an admin turns on, and what to check first

On October 15, 2026, Anthropic moved Code Review out of research preview. It is now available to Claude Team and Enterprise organizations. Code Review is a feature of Claude Code that reads each new pull request (a proposed code change waiting for a teammate to approve it) and posts findings before a human reviewer starts.

What changed from the preview

Anthropic reports three differences:

• Speed. Most reviews now finish about 3x faster, down from roughly 20 minutes to 6–10 minutes.
• Signal. It finds more of the bugs that matter. Each finding is short and states the problem, what triggers it, and the fix.
• Control. Admins set a review depth for each repository: Quick, Standard (the default), or Deep.

How an admin turns it on

1. Enable Code Review in Claude Code settings.
2. Install the GitHub App on your GitHub organization.
3. Select the repositories it should cover.

After that, reviews run automatically on new pull requests. Developers configure nothing.

Choosing depth per repository

The depth setting is the main decision. A reasonable starting split:

• Quick for documentation, internal scripts, and prototypes, where a missed bug costs little.
• Standard for most product code.
• Deep for code that moves money, handles customer data, or controls access. A longer review costs more compute, so keep it for repositories where a bug reaches customers.

Credits for Enterprise

Anthropic says eligible Enterprise organizations can claim at least $10,000 in Code Review credits by January 15, 2027. Credits expire 30 days after they are claimed, and terms apply. The announcement gives no per-review price, so claim the credits when you are ready to run a real pilot, not before. Thirty days is short.

A two-week pilot

1. Pick three repositories: one low-risk, one typical, one critical. Set Quick, Standard, and Deep.
2. For two weeks, have the human reviewer record one line per pull request: did Code Review catch something the reviewer would have missed, did it flag something wrong, or did it add nothing?
3. Count at the end. If wrong findings outnumber useful ones in a repository, lower its depth or remove it.
4. Tell developers that Code Review is an extra reader, and the human approval rule stays as it is. Reviewers who treat a clean automated review as approval are the main risk.

For the wider rollout questions, see Claude Code for operators and Claude enterprise cost controls.