One "@claude review" and multiple agents split the work of reading your PR — Claude Code's Code Review is now generally available
Anthropic has made Claude Code's automated code review generally available for Team and Enterprise. Multiple agents read a PR in parallel, cut false positives, and then surface findings in order of severity. Reviews that once took around 20 minutes now run in 6–10, and we break down the pricing, set
On October 15, 2026, Anthropic moved Claude Code's automated code review feature, "Code Review," from research preview to general availability. It is available to organizations on the Team and Enterprise plans. About seven months after the preview launched in March, the setup in which multiple AI agents read a GitHub pull request (PR) and return their findings as inline comments has become an option for production use.
Out of research preview and into Team / Enterprise
Code Review kicks off automatically when a GitHub PR is opened, among other triggers. It is available on the Team and Enterprise plans, and the repository must be hosted on GitHub. GitHub Enterprise Server, as well as data-residency-enabled GitHub Enterprise Cloud, are also supported.
Reviews run in Anthropic's managed environment. By design, they neither approve PRs nor block merges — the authority to pass or fail stays with the development team. The check run also never ends as a "failure."
Four stages that "cut false positives before commenting"
A review proceeds in the following four stages. This is a summary of the official documentation, not a verbatim reproduction.
- Find: Multiple agents read the diff and surrounding code in parallel.
- Verify: Candidate findings are checked against the code, and false positives are discarded.
- Rank: The remaining findings are deduplicated and ordered by severity.
- Post: Inline comments on the relevant lines, a summary, and a check run are returned to the PR.
During a review, access to the repository is read-only, and no commits are written. It does not build or run the code either — it only reads the diff and the surrounding code. By default, it focuses on correctness issues such as "bugs that break production," leaving formatting preferences and missing tests out of scope.
Three depth levels, triggered by "@claude review"
You can choose from three levels of review depth. The deeper the review, the more time and money it takes.
| Level | Characteristics |
|---|---|
| Quick | A lighter pass. Fastest and cheapest. |
| Standard | Balances depth, speed, and cost (the default). |
| Deep | The most thorough. Takes the most time and money. |
The Owner picks the organization's default level, and it can also be set per repository. For example, you could route routine code to Quick and critical code to Deep. You can also choose when it triggers: "when a PR is opened," "on every push," or "only on request." Under any of these settings, anyone with write access can run a review on the spot by commenting @claude review on the PR.
Reviews that took 20 minutes now take 6–10, and internally findings more than tripled
According to Anthropic, many reviews are now about three times faster than during the preview, with the time required dropping from roughly 20 minutes to around 6–10. Given that waiting on reviews is a common source of PR congestion, this speedup is not something to dismiss.
The company also pointed to the impact of its internal rollout. Before adoption, 16% of PRs received "substantive findings"; after adoption, that rose to 54%. It's Anthropic's own data suggesting the quality and quantity of reviews were both lifted.
Pricing is $15–25 per PR, with up to $50,000 in credits for Enterprise
The cost varies with the size and complexity of the PR and the review level. The preview-era figures reported are as follows.
| Item | Details |
|---|---|
| Guide price per review | About $15–25 (varies with PR size and complexity) |
| Credits for Enterprise | Equivalent to $10,000–$50,000 for eligible organizations |
| Credit application window | October 15, 2026 – January 15, 2027 (23:59 UTC) |
| Validity period | 30 days from application |
Since the cost scales with size, it adds up faster for organizations with more PRs. Code Review also includes a mechanism to set a monthly spending cap, so the practical approach is to decide on a cap before you start running it.
Rollout starts with "the Owner installing the GitHub App"
Enabling the feature is done by the Owner of the Claude organization. The rough flow is: check the requirements, install the Claude GitHub App, and select the target repositories along with the review level and trigger timing. Once configured, reviews can run automatically whenever a new PR is created, without developers having to do anything special.
What gets flagged is tuned via REVIEW.md or CLAUDE.md in the repository. Note that REVIEW.md is read from the base / default branch, not from the PR's own changes. This is a design that closes off loopholes such as rewriting the review instructions on the PR side to slip past inspection.
"No findings" does not mean "safe"
It's worth looking at the flip side of the convenience, too. The official documentation states plainly that "Code Review is an AI-powered review and may miss bugs, security vulnerabilities, and other issues," cautioning that "the absence of findings does not guarantee that the code is flawless or secure." It's best to hold onto the premise that automated review is not a substitute for the final call.
There are also constraints on the deployment environment. It cannot be used by organizations that have applied Zero Data Retention (ZDR) or a HIPAA configuration, and organizations using customer-managed encryption keys (CMEK) are not supported at this time either. It is not covered by Anthropic's Business Associate Agreement (BAA) either. In heavily regulated settings, you'll need to confirm the requirements up front.
Separately from the organization-level general availability, the /code-review command — which reviews your local diff within an individual session — is available on any plan. It makes for a good entry point if you want to start small.
References: Code Review - Claude Code Docs / Unite.AI / Superpower Daily / AlphaSignal
