AI Code Review Tools: Claude Code Review Goes GA, Pricing Shifts

The review layer becomes the main event

For two years, the race between AI coding assistants was about who could write the most code fastest. In 2026 the bottleneck moved to reviewing that code. On October 15, Anthropic made Code Review for Claude Code generally available to Team and Enterprise organizations. The feature is a multi-agent system that checks pull requests automatically, and Anthropic is offering eligible Enterprise customers at least $10,000 in promotional review credits to go with it1. One report puts the top of that credit range at $50,000, with the exact amount set by how many active Claude Code seats an organization has6.

The launch caps a year in which all three major platforms (Anthropic, GitHub and Cursor) rebuilt their review products around agents and moved them to metered pricing. Read together, those changes show where the market is going: review is now a separate, billable step in the software pipeline, not an extra bundled with an editor subscription.

What Anthropic shipped

Code Review first appeared on March 9 as a research preview. Anthropic said its engineers' code output had grown 200% in a year, and that review had become the constraint as a result18. Cat Wu, Claude Code's head of product, described the problem this way: a single prompt from one engineer can produce a plausible-looking pull request, and the reviewer then has to spend real time checking its edge cases2.

According to Anthropic, the general-availability version is about three times faster than the preview. Typical reviews now take 6–10 minutes, down from roughly 2016. Administrators can set each repository to a Quick, Standard or Deep review level, and each finding is now written up as the problem, what triggers it, and the fix1. Under the hood, several agents read the diff and the surrounding code in parallel. A verification pass then filters out false positives, and the remaining findings are deduplicated and ranked by severity before they are posted16. The system leaves inline comments and a check run, but it never approves or blocks a merge6.

The internal numbers Anthropic has published since March are strong. The share of its pull requests that received substantive review comments rose from 16% to 54%. Large pull requests of more than 1,000 lines produced findings 84% of the time, averaging 7.5 issues each. Engineers marked fewer than 1% of findings as incorrect27. Those figures deserve some skepticism. VentureBeat found that "marked incorrect" means an engineer resolved a comment without fixing it, and argued that Anthropic still needs outside validation to justify its premium price8.

There are hard limits. The managed service works only with GitHub. It is not available to organizations that use Zero Data Retention, the HIPAA configuration or customer-managed encryption keys, and it is not covered by Anthropic's Business Associate Agreement16. For a product aimed at large enterprises, that rules out much of healthcare and other heavily regulated industries for now.

GitHub's quiet rebuild of Copilot review

GitHub Copilot news this year has been a steady run of changelog entries rather than one big launch. On March 5, Copilot code review moved to an agentic tool-calling architecture that pulls in broader repository context, and it became generally available on all paid Copilot tiers26. Later updates added support for AGENTS.md files in June22. Agent skills and read-only MCP server connections reached general availability in July21. In August, GitHub removed the old 300-file and 20,000-line size limit and added full reviews of pull requests opened by Copilot's own cloud agent25.

The September updates point in the same direction as Anthropic's. Copilot review can now use shell tools to build code, run tests and execute scripts. Its Lite effort level now runs an ensemble of agents rather than a single one, which GitHub says raised the number of addressed high-severity comments per review by 47% while cutting review cost by about 8%27. A redesigned overview comment sorts findings into open, resolved and previously missed23.

This matters because Copilot review was once seen as the fast but shallow option. Some 2026 comparisons still describe it as diff-based and traded off depth for convenience39. The changelog record suggests that view is out of date: GitHub has adopted the same multi-agent, verify-before-posting design that Anthropic promotes.

Pricing changed too. Since June 1, each Copilot code review on a private repository is billed in AI Credits and also uses GitHub Actions minutes24. In practice, a review that once cost nothing extra on top of a seat now has a running cost.

Cursor's Bugbot and the pricing confusion

Reporting on Cursor's Bugbot contradicts itself more than coverage of any other tool. Several 2026 guides still list it as a $40-per-user monthly add-on161718. Cursor's own documentation tells a different story. Since a May 11 announcement, Bugbot has used usage-based billing with no seat fee, and an average run costs $1.00 to $1.50 depending on the size and complexity of the pull request. Existing subscribers move to the new model at their first renewal after June 81115. A June 10 update reportedly cut review time from about five minutes to roughly 90 seconds and made runs about 22% cheaper1314.

The bigger issue is what counts as a run. Bugbot reviews every update to a pull request by default, so a branch that goes through several rounds of pushes can be billed several times. Cursor has also not published a spending cap12. That comes from a competitor's cost calculator, but another source agrees that each push adds a run, though later runs only review what changed since the previous one14. Coverage also disagrees on platform support. Older reviews call Bugbot GitHub-only1719, while more recent comparisons list GitLab, Bitbucket and Azure DevOps, some still in beta3436. The newer, documentation-based accounts are more likely to be right.

One pricing guide also reports that SpaceX acquired Cursor in August 2026 and that Bugbot is still on sale14. If that is accurate, it adds uncertainty to the long-term roadmap of the Cursor editor.

Why the rankings disagree

Search for the best AI code review tools in 2026 and you will find many ranked lists, most of them published by vendors that rank themselves first. Cubic puts itself at the top of Martian's Code Review Bench with an F1 score of 64.0, and admits it is its own product36. Greptile names itself the best choice for reviewing AI-generated code32. Optimal recommends its Optibot for most teams35, and Qodo calls itself the best option for GitHub38.

The benchmark data that does exist shows a crowded field. On cubic's snapshot of Martian's leaderboard, the top three tools are within 0.8 F1 points of each other. Claude scores 60.9, GitHub Copilot 59.9 and Cursor Bugbot 58.036. Differences that small are unlikely to matter to most teams as much as integration and cost.

The less self-interested guides tend to recommend the same setup: one AI reviewer matched to your Git host, plus a rule-based static-analysis gate such as SonarQube or Snyk, plus ordinary linters31. Kodus's comparison likewise frames the decision around which Git provider you use and how much context a tool can draw on, rather than raw accuracy34.

The real story is cost structure

The clearest trend across all this coverage is that review is becoming metered. Anthropic charges by tokens, averaging $15–25 per review3. Copilot uses AI Credits plus Actions minutes24. Bugbot charges per run11. Independent players such as CodeRabbit and Greptile mostly still sell seats, at around $24 and $30 per developer per month respectively32.

That gives buyers a real choice. A seat-based reviewer gets cheaper per review the more pull requests each developer opens. Usage-based tools scale directly with how much code is shipped, and agent-generated code is pushing that volume up sharply. One estimate shows Bugbot costing more than seat-based competitors at 1,000 pull requests a month, while favoring teams that open fewer pull requests per developer11.

Our reading is that Anthropic has priced Code Review as a premium product for depth on high-stakes changes, not as a reviewer for every pull request. The configurable review levels and the large promotional credits look designed to get enterprises past the sticker shock1. GitHub's advantage is distribution: its reviewer is built into the platform where the code already lives, and it has quickly closed the quality gap. Cursor is using cheaper, faster runs to keep teams that already work in its editor.

For engineering leaders, the practical takeaway is to budget for review as its own cost line, check current vendor pricing documentation rather than third-party lists, and keep a human approving every merge. Every vendor still insists on that last point, and Anthropic's own documentation warns that a review with no findings does not mean the code is secure1.