AI Code Review Tools Move to Per-Review Billing in 2026

AI code review now runs on several agents and bills per review

In 2026, AI code review became one of the main products sold by the companies behind coding assistants. Over a few weeks this autumn, the three biggest names in the category moved in the same direction. Anthropic made Code Review for Claude Code generally available to Team and Enterprise organizations on October 15. It also offered eligible Enterprise customers at least $10,000 in promotional review credits.31 At GitHub Universe, GitHub announced a Max tier for Copilot code review. Max runs several independent review agents built on models from different providers, and an aggregation agent then combines and checks their findings.27 Cursor had already dropped Bugbot's flat seat fee in May and now bills by the review run.1215

The products work in similar ways, and the vendors explain them in similar terms. Several agents review a change in parallel, a verification step removes weak findings, and the bill tracks how much work each review took. The real question for engineering leaders is no longer whether to buy an AI reviewer. It is how much each review should cost and how far its output can be trusted.

Why review became the bottleneck

Vendors and reporters give the same reason for the shift: coding agents produce pull requests faster than people can read them. Anthropic says code output per engineer rose 200% over the past year, and that review became the constraint as a result.31 Cat Wu, Anthropic's head of product for Claude Code, put it this way: one engineer with one prompt can now produce a plausible-looking PR, and the reviewer is left to check every edge case.33 According to Anthropic's own figures, 16% of its internal PRs received substantive review comments before it deployed the tool. Afterward, 54% did.3540

GitHub's Universe announcements make the same point. Analysts quoted on the new Copilot Projects feature warned that ten parallel agent sessions could produce ten pull requests, each needing review, testing and integration. That could mean more merge conflicts and heavier CI load.27 Earlier in the year, GitHub had Copilot's coding agent run Copilot code review on its own changes before opening a PR, so the agent tidies its own work before a human sees it.26 Put simply, review is now the step that decides whether agent-written code can actually ship safely.

Claude Code Review: slower, deeper and more expensive

Anthropic has openly favored depth over speed and price. Each review goes through four stages: parallel agents look for issues, a verification pass filters out false positives, the remaining findings are deduplicated and ranked by severity, and the results are posted to the PR.31 The reviewer never approves or blocks a pull request. Its check run always ends as neutral, so teams that want findings to gate merges have to wire that into their own CI.36

The general-availability release dealt with the biggest complaint from the preview, which was speed. Reviews took about 20 minutes at launch.33 Anthropic says most now finish in 6 to 10 minutes. Administrators can also set each repository to Quick, Standard or Deep review.31 During the preview, reviews averaged $15 to $25 in token usage.3137 That is far more than any competitor charges per review. One comparison estimated that a 20-developer team merging 200 PRs a month would spend about $4,000 a month on Claude reviews, against about $800 for Bugbot's old seat pricing.19

The limits on who can use it matter just as much. The service supports only GitHub-hosted repositories. It is not available to organizations with Zero Data Retention or Anthropic's HIPAA configuration, and it does not run in self-hosted environments.31 Some regulated buyers will be ruled out entirely. There is also some confusion about access. One vendor's explainer says in one place that the feature launched for Pro and Max subscribers, and elsewhere that individual plans cannot use it.32 Anthropic's announcements and independent coverage agree that only Team and Enterprise plans get the managed service.3337 Individual developers have the local /code-review command, which is a separate product.36

GitHub Copilot: frequent updates and tiered effort levels

Copilot's approach is to make review a default part of GitHub, with a series of steady updates. Copilot code review moved to an agentic, tool-calling architecture and became generally available on March 5.25 Later updates added several features:

• support for AGENTS.md files22
• agent skills and read-only MCP connections21
• reviews of very large PRs, with the earlier 300-file or 20,000-line limit removed24
• shell tools that let the reviewer run builds and tests23

GitHub says the September Lite update uses a group of agents instead of one. In its own tests, that increased addressed high-severity comments by 47% and cut review cost by about 8%.23 These are GitHub's internal numbers, not independent results.

The bigger change is pricing. Since June 1, Copilot usage has been billed in AI Credits worth one cent each, and reviews on private repositories also use GitHub Actions minutes.29 GitHub now estimates average costs of $0.25 for a Lite review, $1.50 for Balanced and $10 for the new Max tier.27 One third-party comparison gives wider ranges of $0.05 to $1 for Lite and $0.25 to $5 for Balanced, and notes that Balanced became the default on September 28.9 Teams that budgeted on the assumption that review was included in the Copilot plan should check their numbers again.

GitHub also differs from Anthropic on approvals. Copilot can submit approving reviews, an opt-in preview available since September 1.925 Anthropic's reviewer never approves.33 This analysis sees that as the clearest governance difference between the two. Teams should treat Copilot's auto-approval as a policy decision, not a convenience setting.

Cursor Bugbot: pricing still inconsistent across sources

Of the three, Cursor's reviewer has the most inconsistent coverage. Many comparison guides still list Bugbot as a $40-per-user monthly add-on.1519 Cursor's own documentation, as summarized by several sources, says the product moved to usage billing on May 11, 2026. The average run now costs $1.00 to $1.50, and existing subscribers switch at their first renewal after June 8.1215 One cost-tracking site gives a much lower figure of about $0.27 for a typical 500-line PR after a June efficiency update.13 Other sources do not confirm that number, so buyers should treat it with caution.

The word "run" is the part to watch. One analysis says Cursor staff confirmed the charge is per run, not per PR, and that Bugbot re-reviews on every push by default, so a branch that goes through several pushes is billed several times.11 Another source says Cursor's documentation does not say whether new pushes count as new runs.12 Since the two disagree, teams should test this themselves before rolling Bugbot out widely.

Reports on platform support also differ. Some guides describe Bugbot as GitHub-only or limited to the Cursor ecosystem.1820 Newer comparisons list GitHub, GitLab, Bitbucket and Azure DevOps, with some of those still in beta.917 The newer listings are probably more accurate. Ownership has changed as well: SpaceX acquired Cursor in August 2026, and Bugbot is still on sale.17

Benchmarks are hard to trust

No shared standard exists for comparing these tools. Greptile scored 82% in its own July 2025 test, but 36.1% in a 2026 test by Tenki. Copilot fell from 54% to 24.6% between the same two tests.2 Macroscope's benchmark puts it ahead of Bugbot, at 48% detection against 42%.16 Cursor reports resolution rates, not precision or recall.15 VentureBeat noted that Anthropic has not published a controlled evaluation against a human-established ground truth.40

Most of the published numbers come from vendors measuring their own products or their competitors. A reasonable conclusion is that these tools now find real bugs, but nobody can say with confidence which one finds the most.

How to choose

Independent reviewers such as CodeRabbit, Greptile, Qodo and Graphite still compete on flat seat prices and on support for more Git hosts.9 CodeRabbit is often described as the broadest option across platforms.27 Among the assistant vendors, this analysis sees three distinct positions:

• Anthropic: sells high-cost, high-depth reviews for large and risky changes.36
• GitHub: sells tiered effort levels built into the platform most teams already use.27
• Cursor: sells metered reviews tied to its editor.16

Many teams will probably run more than one. A cheap default reviewer can check every PR, with a more expensive deep pass reserved for code paths where a missed bug would cost far more than the review.