Claude Code gets AI code review in beta

What's new

Anthropic has launched a beta version (a research preview) of a new feature in Claude Code called Code Review. The feature automatically kicks off multiple AI agents every time a developer opens a pull request, meaning a code change that must be approved before it goes into the actual program. The agents review the code in parallel, look for bugs, double-check each other's findings to avoid false alarms, and then rank the problems by severity. The result is shown as one overall comment plus precise comments directly in the code.

What's clever

The clever part is that the system itself judges how closely to scrutinize. A small change of under 50 lines of code gets a quick review, while a large change of more than 1,000 lines puts more agents to work and digs deeper. At Anthropic itself, it has raised the share of pull requests that get a thorough comment from 16 percent to 54 percent. On large changes, the system finds problems in 84 percent of cases, with an average of 7.5 bugs per review. Most importantly: fewer than 1 percent of the system's findings are rejected by developers as wrong, which is unusually low for this type of automated tool.

Cheaper or better?

This is no free ride. A review costs between $15 and $25, corresponding to around 3 million tokens (the chunks of text AI models compute in) on Anthropic's Opus model. That makes it more expensive than lighter competitors such as GitHub Copilot's built-in code review and CodeRabbit, but Anthropic is selling the feature precisely on being more thorough rather than fast and cheap. It is only available to companies on the Team and Enterprise plans, so not to regular Pro or free users. Nothing specific has been said about Denmark or the EU, but the feature is available globally to the customer types mentioned.

What it's good at

The tool is made for companies and developer teams that want to catch bugs before they become part of the finished product, especially in large and complex code changes where human eyes often miss something. Companies can set a monthly budget and control which projects use the feature, so costs don't spiral out of control. For the individual developer, it means in practice an extra, thorough proofreader that works in the background and is typically done in about 20 minutes.