Anthropic launched Code Review on Monday inside Claude Code, a tool that reads pull requests and flags bugs and logic errors in the growing pile of AI-generated code. It arrived in research preview for Claude for Teams and Claude for Enterprise, and at an estimated $15 to $25 per review, it treats verification as a premium service rather than an afterthought.
The launch says something bigger than one product. When machines write code faster than people can read it, the scarce resource stops being generation and becomes review. That same inversion is coming for how brands show up inside AI answers, which is why this matters even if you never touch a codebase.
Key takeaways
- Code Review runs in Claude Code and integrates with GitHub, analyzing pull requests automatically and leaving inline comments that explain the issue, why it is risky, and how to fix it. - It uses a multi-agent architecture: several agents examine the same code in parallel from different angles, then a final pass deduplicates and ranks findings by severity, using color tags with red as the highest. - The tool deliberately prioritizes logic errors over style, because head of product Cat Wu says earlier AI feedback tended to be noise developers could not act on. - At roughly $15 to $25 per review, verification is now a paid, resource-intensive layer, a sign that checking AI output is becoming its own category of work. - The GEO parallel is direct: as AI generates more of what represents your brand, from product copy to structured data, the constraint shifts from producing content to verifying that machines read it correctly.
What Code Review actually does
The problem Anthropic is naming is one its own tools helped create. Vibe coding, where you describe intent in plain language and the model emits large amounts of code, has sped up development and introduced new bugs, security gaps, and code nobody fully understands.
Wu framed it through a question enterprise leaders kept asking: now that Claude Code is opening a flood of pull requests, how do you get them reviewed efficiently? Pull requests are the checkpoint where a human signs off before changes reach the main codebase. When one AI dramatically increases how much code gets written, the review queue becomes the choke point on shipping.
Code Review is the answer to its own side effect. Rather than nitpicking formatting, it focuses on logic, the class of error most likely to cause real damage and least likely to be caught by a linter. The multi-agent design matters here: a single reviewer, human or model, sees code from one vantage point, while parallel agents each probe a different failure mode, and the aggregation step keeps the output short enough to act on.
Why "review" is the real story
For two years the industry optimized for output. The interesting shift now is that trust, not speed, is the binding constraint. Anthropic pricing a review at $15 to $25 is an admission that carefully checking machine-written work is expensive and that people will pay for it because the alternative is worse.
