Claude Code Review leaves preview and runs 3x faster

"Coding agents keep increasing the pace at which teams write code, and the review bottleneck hasn't eased," Anthropic writes in its general availability announcement on Claude. Code Review, Anthropic's tool for that bottleneck, has left research preview and is available to Claude Team and Enterprise organizations. It takes the first pass on every new pull request automatically.

Anthropic's account: faster reviews, shorter findings and a level for each repository

Anthropic says the GA version is better at catching the bugs that would otherwise reach production. It also says each finding is now short: a few sentences that state the problem, what triggers it and the fix. The company runs Code Review on its own pull requests and says this shows where reviews slow its engineers down.

The rest came from preview customers. According to Anthropic, they said the tool caught real issues before a human reviewer opened the PR, but they wanted reviews that were quieter, sized to each change, and faster. The three review levels are aimed at that request. Routine codebases get a quick pass and critical ones get a deep review. Anthropic says every level delivers high-quality reviews, but it does not explain what Quick skips that Deep covers.

Admins set it up in three steps: enable Code Review in Claude Code settings, install the GitHub App and pick the repositories. Developers configure nothing. Reviews just start appearing on new PRs.

What the preview-era sources add

According to TechCrunch, the preview launched on Monday, March 9, 2026, for Team and Enterprise customers and was aimed at large enterprise Claude Code users. Cat Wu, head of product for Claude Code, named Uber, Salesforce and Accenture. Speaking to The New Stack, Wu described the problem as "it only takes one engineer, one prompt, to put out a plausible-looking PR." The reviewer then has to verify the edge cases.

Both outlets reported the same design choice. Code Review targets logic errors rather than style, to keep false positives down. "People are very sensitive to false positives," Wu said, and a real bug is something you "should almost definitely fix."

Anthropic's preview announcement on Claude gave the numbers that the GA post leaves out. That post says code output per Anthropic engineer grew 200% in a year, and that many PRs "get skims rather than deep reads." Here is how it describes the mechanism. When a PR opens, a team of agents looks for bugs in parallel, verifies each candidate to filter out false positives, ranks the survivors by severity, and posts one overview comment plus inline comments. Larger changes get more agents.

The same post reports these internal figures. Before Code Review, 16% of Anthropic's PRs got substantive review comments, and with it 54% did. On PRs over 1,000 lines, 84% get findings, averaging 7.5 issues. Less than 1% of findings are marked incorrect. Preview reviews were billed on token usage and "generally average $15–25," scaling with PR size and complexity. Admins had monthly organization caps, repository-level controls and an analytics dashboard. Anthropic positioned the product as more thorough and more expensive than the Claude Code GitHub Action, which remains open source and available.

Unite.AI's reading of the docs adds the limits

According to Unite.AI's summary of the documentation, each review runs in four stages: parallel examination of the diff and surrounding code, verification against the code, deduplication and severity ranking, and posting. The check run, named Claude Code Review, never concludes as a failure. A review never approves or blocks a PR, so it cannot gate a merge.

Unite.AI also lists the limits. Reviews run on Anthropic-managed infrastructure with read-only repository access, and they read code without building or running it. They are not available in self-hosted environments or to organizations with Zero Data Retention or the HIPAA configuration. Repositories must be hosted on GitHub. Draft PRs and forks are excluded from automatic reviews. Admins cannot choose the model, and more thorough levels cost more.

Per the same summary, REVIEW.md and CLAUDE.md files tune what gets flagged. REVIEW.md is read from the base or default branch, so a PR cannot loosen its own review rules. Reviews can run once when a PR opens, after every push, or when anyone with write access comments @claude review.

What we don't know

• How well the GA version detects bugs. The 16%-to-54% and under-1% figures are Anthropic's internal preview-era numbers, and the GA post gives no new ones.
• GA pricing. The announcement doesn't mention it. Preview reviews averaged $15–25, and Unite.AI says deeper levels cost more, but no source gives a price for each level.
• Whether the 6–10 minute figure holds for Deep reviews or mainly reflects Quick and Standard ones. Anthropic only says "most reviews."
• Which Enterprise organizations count as eligible for the credits. The terms are referenced but not summarized.

On the evidence we have, the speed claim is the better-supported one, because it comes with a before and an after in minutes. The quality claim does not. Anthropic's preview launch came with percentages, while the GA post says only that the tool catches "more of the bugs that matter." Until Anthropic publishes an updated detection or false-positive rate, the most practical test is to run Code Review on a repository where you already know where the bugs are.