AI Code Review Tools: Claude, Copilot and Bugbot Diverge on Cost

A small test that shows how the market looks

A new hands-on comparison of eight AI code review tools found Qodo, Claude Code and Bito each catching all four production-style defects planted in a single TypeScript pull request16. GitHub Copilot's repository agent also found all four bugs, but it added two questionable extra findings. CodeRabbit and Greptile each caught three, and Graphite caught none16. The planted defects covered authorization, a business rule, payment logic and a bug that spans files. Qodo was the only tool credited with linking a pricing-plan change to limit logic in a file the pull request never touched16.

One pull request with four bugs is a very small sample, and nobody should treat the result as a final ranking. The test is still useful because it captures where this category stands in late 2026. The tools that read a whole repository and run agents are clearly ahead of the ones that only comment on the changed lines. The bigger story, though, is the change in pricing that came with that shift. Anthropic, GitHub and Cursor have all moved their reviewers to usage-based billing, so the cost of AI review now depends on how much code a team pushes, not on how many people it employs.

Claude Code bets on slow, expensive and thorough

Anthropic launched Code Review for Claude Code on March 9, 2026. When a pull request opens, the system sends out a group of agents to look for bugs4. They work in parallel, try to verify each finding to weed out false positives, rank issues by severity, and then post one summary comment plus inline flags1. Anthropic says it never approves pull requests, and that the merge decision still belongs to a person9.

The company's internal numbers are what drive the hype. Anthropic says the share of its pull requests that received substantive review comments rose from 16% to 54% after adoption. Pull requests over 1,000 lines produced findings 84% of the time, averaging 7.5 issues each, and engineers marked fewer than 1% of findings as incorrect5. The trade-off is speed and cost. Reviews take about 20 minutes on average and usually cost $15 to $25, billed by token usage6. Claude Code product head Cat Wu told The New Stack that the longer run time is deliberate, because each agent can move through the whole codebase and not just the diff5.

Coverage disagrees on who can use it. Anthropic's announcement and most reporting say the managed product is a research preview limited to Team and Enterprise plans92. Some third-party guides suggest review is folded into normal Claude Code usage27. The likely explanation is that "Claude Code review" now refers to three different products. One is the managed GitHub reviewer. Another is a local /code-review command that runs on ordinary Claude Code usage. The third is an open-source GitHub Action that runs on a team's own runners8. Buyers should check which one a vendor comparison is actually describing.

One ranking puts Claude Code Review first for correctness on large pull requests. That placement leans heavily on Anthropic's own figures, which no independent party has measured13. Developer reaction at launch was mostly positive about the depth of review, with doubts about whether $15 to $25 per pull request works for teams that ship a lot of code1.

Copilot turns review into a metered agent

GitHub has spent 2026 rebuilding Copilot code review piece by piece. In March it moved to an agentic, tool-calling design that pulls in wider repository context. That design runs on GitHub Actions. From June 1, each review is charged as AI Credits and also uses Actions minutes on private repositories31. For teams that thought of Copilot review as an included feature, this is a real change.

The feature updates have come about once a month. In June, GitHub added organization-level runner controls, support for content exclusions, and removed the character limit on custom instructions34. Agent skills and MCP server connections became generally available in late July, with MCP calls limited to read-only35. In August, the Lite and Balanced effort levels became generally available33. In September, GitHub said Copilot now resolves its own comments once they are fixed and validates code with a full set of shell tools. GitHub also said an ensemble of agents in Lite reviews increased addressed high-severity comments by 47% while cutting review cost by about 8%36. On October 2, GitHub added API-triggered reviews and made Balanced the default effort level32.

In effect, GitHub has adopted much of Anthropic's design: several agents, adjustable depth and repository context. Its advantage is that it is already built into the platform. In the TechnBrains test, Copilot caught all four bugs, but with extra noise16. Not every team is convinced. One publication removed mandatory Copilot review from its main repository rules in July, calling the comments advisory and noisy38.

Cursor's Bugbot moves to per-run billing

Cursor's Bugbot has gone through the biggest pricing change of the three. It used to cost $40 per user per month24. Cursor announced a move to per-run billing on May 11, and existing subscribers switch at their first renewal after June 8. The average run costs $1.00 to $1.50 depending on pull request size26. Cursor also said in June that Bugbot was more than three times faster, 22% cheaper and found 10% more bugs per review, crediting its in-house Composer 2.5 model22.

The main budgeting problem is the billing unit. Bugbot charges per run, not per pull request, and it re-reviews on every push by default. A branch that goes through several rounds of changes is billed several times, and Cursor has not published a spending cap23. Rivals have used this point against it, so their framing is self-serving, but the basic arithmetic is easy to check.

Coverage disagrees on platform support. Several reviews describe Bugbot as GitHub-only2521. A more recent comparison lists GitHub, GitLab, Bitbucket and Azure DevOps, with some in beta22. Fast-moving product changes are the likely reason, so dated guides should be treated as stale. The corporate picture has changed too. Graphite is reported to be owned by Cursor19, and one competitor reports that SpaceX acquired Cursor in August28. That second claim appeared in only one place in this coverage.

The benchmark problem

The hardest part of this market is that almost every ranking comes from someone with a product to sell. One analysis spells this out. Greptile's July 2025 test scored Greptile at 82% and CodeRabbit at 44%. Tenki's 2026 test put Greptile at 36.1% and CodeRabbit at 28.7%. Copilot scored 54% in one test and 24.6% in the other13. The differences are too large to come from method alone. Each vendor picks the bug sets that suit its own tool.

The Martian online tracker is the closest thing to a neutral scoreboard, and even its figures mostly circulate through vendor blogs. According to cubic, Copilot code review ranked fourth at 61.0% F1, the Claude GitHub App sixth at 60.2%, and Bugbot tenth at 57.1%. Cubic ranked itself first822. CodeAnt, which also ranks itself first in its own guide, cites Martian placements too11. When independent tests do appear, they lean qualitative. Monterail tested Copilot, Bugbot and CodeRabbit on an internal project and chose CodeRabbit, mostly for its pull request summaries rather than raw bug detection15.

The takeaway is that a published catch rate is a claim, not a measurement. The TechnBrains approach of planting known defects and running every tool on the same code is the more honest method. Teams can repeat it on their own repositories at little cost, and several guides recommend a two-week trial on live pull requests for that reason19.

What it means for engineering teams

The pricing shift matters most in practice. Claude Code Review at $15 to $25 a run is a premium option for large pull requests where a missed bug is expensive19. Bugbot at roughly $1 to $1.50 per run is cheap per review but hard to predict over a month23. Copilot now combines AI Credits with Actions minutes31. Flat per-seat pricing is still offered by tools like CodeRabbit at $24 per user and Qodo from about $3011, and smaller rivals are leaning on predictable bills as a selling point.

The agentic tools do find more real bugs, especially the logic and cross-file errors that human reviewers who skim tend to miss. Anthropic's own reasoning still holds: AI agents produce more pull requests, so review becomes the bottleneck10. The open question is no longer whether to automate review. Teams now have to decide which pull requests justify a $20 deep review and which only need a $1 check. Anthropic built the slow, deep option. GitHub now offers both through effort levels. Cursor is competing on speed and per-run price. Effort levels and per-run billing suggest the market is heading toward routing each pull request to the depth it needs.