AI Code Review Tools Shift to Usage Pricing as Claude Hits GA

A crowded category gets a benchmark, and a price war

A new roundup ranks the ten best AI code review tools of 2026 using a benchmark of roughly 300,000 pull requests. It landed as the three biggest AI coding platforms are changing how their own reviewers work and what they cost. The ranking comes from CodeAnt AI, and it relies on the Martian Code Review Bench. The guide says researchers from DeepMind, Anthropic and Meta built that benchmark and open-sourced it in February 2026.21 Unsurprisingly, CodeAnt names itself the best overall platform. It credits CodeRabbit with the fastest PR summaries, Greptile with the deepest analysis of multi-service architectures, and SonarQube with the most dependable compliance tooling.21

The list matters less than what happened around it. Within a few weeks, Anthropic took Claude Code's Code Review out of preview, GitHub added API access and a new default depth to Copilot's reviewer, and Cursor finished moving Bugbot from seat licenses to per-run billing.13112 Reviewing code used to be a feature bundled with a subscription. It is quickly becoming a metered service of its own, and that changes how teams should judge any "best tools" list.

Anthropic's Code Review goes general availability

On October 15, 2026, Anthropic made its multi-agent Code Review available to all Claude Team and Enterprise organizations.13 The company says most reviews now take 6–10 minutes, about three times faster than the roughly 20 minutes they took during the preview.310 Admins can also set each repository to Quick, Standard or Deep review, with Standard as the default.10 Anthropic says findings now explain the problem, what triggers it, and the fix in a few sentences.1

The design hasn't changed. Several agents read the diff and the code around it in parallel, a verification step weeds out false positives, and the remaining findings are deduplicated, ranked by severity and posted to the pull request.1 The reviewer leaves inline comments and a check run, but it does not approve or block merges.6 Teams can steer it with REVIEW.md and CLAUDE.md files.6

When it launched in March, Anthropic said the share of its internal PRs getting substantive review comments rose from 16% to 54% after rollout. It also said engineers marked fewer than 1% of findings as incorrect.7 On PRs over 1,000 lines, 84% received findings, averaging 7.5 issues each.7 These are the company's own figures from its own codebase, and they should be read that way.

Cost is the main catch. Preview reviews were billed by token usage and averaged $15–25 each.46 When InfoQ covered the preview, some developers questioned whether that price and the 20-minute wait would hold up in high-volume workflows.2 General availability addresses the speed concern but not the price. Instead, Anthropic is offering promotional credits. Here the coverage differs: Anthropic and several outlets describe at least $10,000 for eligible Enterprise organizations, while one report says the amount ranges from $10,000 to $50,000 depending on active Claude Code seats.36 Both versions agree that the credits expire 30 days after they're claimed and that the claim window closes January 15, 2027.110

The product's limits are just as relevant to buyers. It works only with GitHub. It isn't available to organizations using Zero Data Retention, the HIPAA configuration or customer-managed encryption keys, and it isn't covered by Anthropic's Business Associate Agreement.1 That rules out a large share of regulated buyers. Secondary coverage is also inconsistent: one vendor explainer states in one place that the tool launched for Pro and Max subscribers, then says individual plans have no access.5 Anthropic's own materials confirm that only Team and Enterprise plans have it.3

GitHub Copilot turns review into a configurable, metered agent

GitHub has been rebuilding Copilot code review all year. A March 2026 overhaul switched it from line-by-line diff comments to an agentic, tool-calling design that explores the wider repository.38 Because that design runs on GitHub Actions, GitHub began charging for each review twice from June 1: AI Credits for model usage, plus Actions minutes for reviews on private repositories.32 Public repositories still don't use Actions minutes.32

Over the summer, GitHub added support for organization-level runner settings and content exclusions, then for AGENTS.md.3639 In July, Copilot started reading REVIEW.md, GEMINI.md and CLAUDE.md files and moved reviews behind a firewall by default.33 Copilot's reviewer now reads Anthropic's configuration file, which shows how tools are converging on the same repository conventions. In August, the Lite and Balanced effort levels became generally available.40 On September 28, Balanced became the default, and since October 2 teams can request reviews through the REST and GraphQL APIs.31

The Balanced default is worth noting. It sends reviews to a higher-reasoning model, which in practice means more compute per review on a bill that is now metered.4032 Copilot's appeal used to be that review was effectively free for existing subscribers. That argument is getting weaker, though Copilot still has the lowest barrier to adoption of any tool here.30

Cursor's Bugbot drops the $40 seat

Cursor announced on May 11, 2026 that Bugbot would move from a $40-per-seat subscription to usage billing. Existing subscribers switch at their first renewal after June 8.1612 Cursor now says an average review costs $1.00–$1.50, depending on PR size and complexity.11 According to one analysis, Cursor said in June that Bugbot had become 22% cheaper. The same analysis reports that SpaceX acquired Cursor in August and that Bugbot is still on sale.14

What counts as a "run" matters. Critics point out that Bugbot reviews every PR update by default, so a branch that goes through several rounds of revisions is billed several times. They also note that Cursor hasn't published a spending cap.13 One cost analysis found the new model cheaper than the old seat plan for teams with modest PR volume. A 10-person team doing 2,000 reviews a month, though, would go from $400 to roughly $2,000–$3,000.12

The CodeAnt ranking still lists Bugbot as a $40-a-month add-on available only in the Cursor IDE.21 That reflects the old pricing, and another comparison shows Bugbot supporting GitHub, GitLab, Bitbucket and Azure DevOps, some of them in beta.29 Treat older roundups' pricing columns with caution.

Benchmarks that disagree with each other

The ranking's 300,000-PR claim sounds authoritative, but benchmarks in this category often contradict one another. Greptile's own test credits Greptile with 82% and CodeRabbit with 44%. Tenki's 2026 test puts Greptile at 36.1% and ranks its own product first.22 DeepSource's harness puts DeepSource on top with 84.5% F1 and gives CodeRabbit 36.0%.26 On the Martian leaderboard's last-month view as of October 10, cubic leads at 64.0 F1, followed by Kody AI at 63.4 and Greptile at 63.2, with Claude at 60.9, Copilot at 59.9 and Bugbot at 58.0.29 cubic publishes that ranking and says so.29 CodeAnt, for its part, ranks itself third on Martian at 51.7%.21

Two conclusions follow. First, almost every vendor ranks first on a benchmark it chose or built. Second, on the most independent leaderboard available, the leading tools are separated by only a few F1 points. The roughly six-point spread between cubic and Bugbot is small compared with the gaps in pricing, platform support and data-handling terms. Where independent reviewers agree is on the shape of the market. CodeRabbit is repeatedly described as the low-noise all-rounder that supports all four major Git hosts, and Greptile as the tool that catches the most at the cost of more false positives.2627

What it means for buyers

In our view, the main development of 2026 isn't any single tool winning. It's that the three companies whose agents produce most AI-generated pull requests now sell the reviewers that check them, and all three charge per review. Anthropic's rationale is explicit: it says code output per engineer grew 200% and review became the bottleneck.1 The pricing reflects how thorough each tool tries to be. Anthropic's deep reviews cost $15–25, Bugbot runs cost about $1–1.50, and Copilot charges credits plus Actions minutes.41132

Several guides recommend a layered setup: an AI reviewer for logic, a SAST scanner for vulnerabilities, and a deterministic gate such as SonarQube for auditable standards.26 That still looks like the safest approach. Teams should also check data-retention eligibility, platform support and pricing that scales with every push before trusting any leaderboard. On the independent benchmark, the gaps between tools are within a few points. Bills can differ by a factor of ten.