Anthropic debuts extremely efficient but pricey code checking tool for developers

When it comes to writing software, getting feedback is a critical part of the process, ensuring that bugs in the newly written code can be caught early, before a pull request is submitted. But with the rise of artificial intelligence coding bots, developers are shipping more code than ever before, overwhelming human reviewers.

Anthropic PBC has come up with a solution, announcing today the availability of Code Review in Claude Code, a new multi-agent system that’s designed to spot bugs in AI-generated code before a human reviewer ever sees it.

The new product is meant to review pull requests, which are a mechanism used by developers to submit code changes for review before they’re implemented in the software. With most developers using tools like Claude Code to accelerate their output, there has been a dramatic increase in the number of pull requests at many organizations, creating a new bottleneck in software development.

The launch of Code Review comes at a critical juncture in Anthropic’s story. Earlier today, the company filed two lawsuits against the U.S. Department of Defense after it was designated as a supply chain risk. The designation threatens to derail Anthropic’s booming government business, so it makes sense that the company wants to double down on its enterprise customer base, where subscriptions have quadrupled since the beginning of the year.

Claude Code is the company’s most popular enterprise product, and its annual revenue run-rate recently surpassed $2.5 billion. It’s hoped that Code Review will help to make its coding tool even more attractive.

Anthropic has previously integrated code checking capabilities within Claude Code, giving it the ability to review its own work. In addition, there’s a Claude Code GitHub Action tool that can be set up to automatically review code as part of a company’s continuous integration/continuous development pipeline. But Code Review is meant to go further and conduct more comprehensive reviews, although companies will have to pay a high price for the privilege.

Code Review notably takes much more time to review each pull request, and as it works its way through the code it will explain its reasoning step-by-step, Anthropic said. For each potential bug it finds, it will outline what the issue is and explain why it’s likely to be problematic. It will also offer a recommended fix for that issue. It will label each problem it surfaces according to its severity, with red used for the most severe issues, yellow for possible problems that need review, and purple for issues tied to pre-existing code and historical bugs.



According to Anthropic, Code Review does this by relying on multiple AI agents that work in parallel, with each one examining the codebase from a different perspective. Once that’s done, another agent will come in to aggregate and rank the findings of those agents, remove any duplicate issues and prioritize them based on their order of importance.

It’s extremely comprehensive, but that kind of attention to detail doesn’t come cheap. With so many agents involved in the process, customers are going to burn through a lot of tokens. “Reviews are billed on token usage and generally average $15 to $25, scaling with pull request size and complexity,” the company said. That’s definitely not cheap. In contrast, a service such as Code Rabbit, which also uses AI to review pull requests, charges $24 per month.

Code Review is also on the slow side. Anthropic said the time it takes to review each pull request will vary, but averages around 20 minutes to complete.

Despite this, Anthropic promised that customers will be delighted with the results. It said it has been using the tool internally for several months, and claims that large pull requests of more than 1,000 changed lines, 84% of its reviews found something of note, and about 7.5 issues on average. For smaller pull requests of less than 50 lines, 31% were flagged with an average of 0.5 issues found.

The company said it has caught some significant bugs too. In one instance involving internal code, the tool flagged a single, innocuous-looking change to a production service that would have disrupted its authentication mechanism.

Code Review is available now in research preview for Claude Code Team and Enterprise subscribers.

Image: Anthropic