Anthropic Ships Claude Code Review — Parallel Agents on Every PR at $15–$25

Anthropic launched Code Review for Claude Code last week, and the architecture is worth understanding. Rather than a single-pass review, it dispatches a parallel team of agents on every PR — one to hunt for bugs, others to validate false positives, a final agent to rank findings by severity. The scale of the review adapts to the size of the PR.

How It Actually Works

When a PR is opened, Claude Code Review:

1. Spawns parallel agents proportional to the diff size — small PRs get fewer agents, large PRs (1,000+ lines) get more
2. Each agent focuses on a specific concern — logic errors, security issues, test coverage gaps, API contract violations
3. A validation agent cross-checks findings before surfacing them — this is what drives the less-than-1% false-positive disagreement rate
4. A ranking agent prioritizes by severity and actionability

The result on large PRs: an average of 7.5 findings with an 84% finding rate. That’s not a single model reading your diff top-to-bottom — it’s a structured review pipeline.

REVIEW.md — The Customization Hook

The most interesting developer-facing feature is REVIEW.md. Drop this file in your repository root and Claude Code Review uses it as its review charter — which rules to enforce, which patterns to flag, which concerns to ignore.

This solves a real problem with generic AI code review: it doesn’t know your team’s domain-specific conventions. A financial services team can tell it to always flag unvalidated decimal arithmetic. An API team can tell it to enforce versioning conventions. A security-focused team can define their threat model.

This is how you make the tool actually useful for your codebase rather than generic.

The Cost Model

$15–$25 per review with optional monthly caps. That pricing is worth benchmarking against your current approach:

• A senior engineer spending 30–45 minutes on a meaningful PR review has a real opportunity cost
• Static analysis tools (SonarQube, Semgrep) catch syntax and pattern issues but miss logical errors
• The $15–$25 range slots Claude Code Review as a supplement to, not replacement for, human review — catching the obvious errors so human attention can go to architecture and intent

The monthly cap option matters for teams with high PR volume. Without it, a busy week could generate unexpected costs.

What It Doesn’t Replace

Human code review for architectural decisions, product tradeoffs, and domain knowledge stays essential. Claude Code Review is optimized for finding bugs and security issues — the things a tired reviewer misses at 5pm on a Friday, not the things that require understanding business context.

Currently in research preview for Team and Enterprise Claude Code subscribers. Access at claude.com/code-review.

Source: Anthropic — claude.com/blog/code-review | WinBuzzer, March 10, 2026