Anthropic has moved Code Review, the multi-agent pull-request review feature of Claude Code, out of research preview and into general availability for Claude Team and Enterprise organizations. According to Unite.ai, the launch happened on October 15, 2026, and came with a promotional credit program aimed at Enterprise customers.

What changed since the preview

Anthropic says it used the preview period to make the system faster, more precise about which bugs it surfaces, and more adjustable per repository. The company reports that most reviews now finish about three times faster, with typical durations dropping from around 20 minutes to between 6 and 10 minutes. Findings are now written to explain the problem, the condition that triggers it, and a suggested fix in a few sentences. Administrators can also assign each repository a Quick, Standard, or Deep review level, so that routine code gets a light pass while critical code gets deeper scrutiny. Preview users reportedly said the tool caught genuine issues before human reviewers looked at a change, but asked for quieter output, effort scaled to the size of the change, and better speed. Anthropic also runs the tool on its own pull requests, which helped reveal where reviews slowed its engineers.

Enabling the feature is straightforward: an administrator switches it on in Claude Code settings, installs the Claude GitHub App, and picks repositories. After that, reviews run automatically on new pull requests without developers needing to configure anything.

How the system works

Per Anthropic's documentation, each review runs on Anthropic's own infrastructure in four stages. Multiple agents analyze the diff and surrounding code in parallel to locate potential problems. A verification stage then tests candidate findings against the code to remove false positives. The surviving findings are deduplicated and ranked by severity, and finally posted to the pull request as inline comments, a summary, and a check run named Claude Code Review. The check run never reports a failure and the tool never approves or blocks a merge, leaving the decision with the human team.

By default the reviewer concentrates on correctness, meaning bugs likely to break production, instead of style preferences or test coverage gaps. Teams can steer what is flagged using REVIEW.md and CLAUDE.md files, and REVIEW.md is read from the base or default branch rather than the pull request's own changes, which prevents a change from altering its own review rules. Anthropic chooses the underlying Claude models; administrators cannot select one, and deeper review levels cost more.

Repositories can trigger reviews when a pull request opens, after each push, or only on demand, and anyone with write access can request one by commenting @claude review. Drafts and forks are excluded from automatic reviews. The reviewer operates in a managed environment with read-only access and analyzes code by reading it rather than building or executing it. It does not run in self-hosted environments.

Requirements and limits

The service needs a Team or Enterprise plan and GitHub-hosted repositories, including GitHub Enterprise Server and GitHub Enterprise Cloud with data residency. It is not available to organizations with Zero Data Retention or the HIPAA configuration, is outside Anthropic's Business Associate Agreement, and does not yet support customer-managed encryption keys. A Claude organization is needed even for developers who reach Claude Code through Amazon Bedrock, Google Cloud's Agent Platform, or Microsoft Foundry. Anthropic cautions that the AI can miss bugs and vulnerabilities, and that a clean review does not prove code is correct or secure.

Internal results from the preview

Anthropic first introduced Code Review as a research preview on March 9, 2026. It said code output per engineer had grown 200% over the prior year, making review a bottleneck. Before the tool, 16% of its pull requests drew substantive review comments; afterward the figure was 54%. For pull requests exceeding 1,000 changed lines, 84% received findings averaging 7.5 issues, while those under 50 lines saw findings on 31% with an average of 0.5 issues. Engineers marked under 1% of findings as incorrect. Anthropic also cited an early-access case in which the tool surfaced a latent type-mismatch bug in TrueNAS's open-source middleware that was silently clearing a ZFS encryption key cache on every sync.

During the preview, reviews were billed by token usage and typically cost $15 to $25 each, scaling with size and complexity. Administrators had monthly organization caps, repository-level controls, and an analytics dashboard covering reviewed pull requests, acceptance rate, and cost.

The promotional credits

Eligible paid Claude Enterprise organizations can claim between $10,000 and $50,000 in credits, with the amount determined at claim time by active Claude Code seats. An Owner or Primary Owner makes a single, irreversible claim in organization settings, which starts a 30-day window during which reviews draw on the credits first. The window ends when 30 days pass or the credits are exhausted. Claims are open until 23:59 UTC on January 15, 2027, and Anthropic's help center advises configuring Code Review on repositories before claiming, since the clock runs regardless of usage.

Exclusions include reseller-billed organizations, those using GitHub Enterprise Server or data-residency instances, trials and lapsed subscriptions, Team, Pro, Max, and Free plans, organizations that took an earlier free Code Review offer, and any organization that cannot use the feature. Failed reviews caused by Anthropic-side problems are not billed or counted, while reviews cancelled by a new push or by the organization count for the cost incurred. Credits do not cover the self-hosted GitHub Action, the /code-review or /ultrareview commands, or other Claude usage.

Analysis

The launch reflects a broader dynamic in AI-assisted development: as coding agents increase the volume of generated code, review capacity becomes the constraint, and vendors are building agentic review to close the gap. Anthropic's decision to keep the tool advisory, never blocking merges, and its emphasis on false-positive filtering signal a focus on trust and developer acceptance. The credit promotion is a clear adoption lever for Enterprise accounts, though the restrictions on GitHub-only support, data retention settings, and self-hosting limit the addressable market, leaving room for competitors such as GitHub Copilot's review features and other code-review tools.