Anthropic announces 'Code Review' to automatically detect bugs in pull requests, with token usage starting from an average of 2400 yen per request



Anthropic has announced 'Code Review,' an advanced multi-agent review system that claims to be able to detect bugs that even human reviewers often miss.

Code Review for Claude Code | Claude

https://claude.com/blog/code-review




Code Review deploys a team of AI agents to pull requests to find bugs that might otherwise be missed. The agents search for bugs, perform validation to reduce false positives, and rank them based on severity. The results are posted as a single summary comment on the pull request, with inline comments added for more specific bugs. The decision to approve the pull request is then up to the user.

The size of the review will vary depending on the PR's content: larger or more complex changes will be assigned more agents and undergo a more in-depth review, while smaller changes will undergo a lighter review.



Anthropic had been using Code Review internally for several months prior to the announcement, and found that it detected issues in 84% of large pull requests (over 1,000 lines of code) and found an average of 7.5 issues. On the other hand, for smaller pull requests (under 50 lines of code), the detection rate was 31%, with an average of 0.5 issues. Most engineers agreed with the findings, and less than 1% of the time, they found errors. In tests, the average review time was about 20 minutes.

According to Anthropic, the amount of code output per engineer at the company has increased by 200% over the past year, which has caused code reviews to become a bottleneck. They've also heard similar stories from customers. This led to a need for reviewers they could trust with all pull requests, and Code Review was born.

Anthropic describes Code Review as 'designed for depth, not speed.' The cost is high because of the thoroughness of the review, and typically costs between $15 and $25 per review.

An OpenAI employee noticed the high price and said, 'If you don't want to pay $25 per review, use OpenAI's products.'




Code Review is available as a beta research preview for Team and Enterprise plans.

Related Posts:

in AI, Posted by log1p_kr