AI Bench Index

AI code review

Compare AI pull-request review tools

Compare documented pricing, review limits, supported Git hosts, deployment options and merge controls. Vendor feature claims are separated from independent evidence of review quality.

Catalogue review recorded · individual claims have separate datesReferenceSourceMethodology

Latest recorded profile review .

Vendor-reported plans and limits. AI Bench Index has not run a first-party code-review harness on these products.

These products comment on pull requests. They are not coding harnesses. Compare Cursor, Claude Code, and Copilot benchmark scores, then come back here if you still need a dedicated reviewer. They also sit on top of model APIs — for the foundation models themselves see our models directory.

Beyond a feature checklist

What a useful code review should establish

A comment count is not a quality score. A substantive finding identifies a concrete defect, explains a reachable failure and proposes a correction that preserves valid behaviour. These original, synthetic examples have local regression tests; no vendor has been evaluated on them.

A real issue needs a failing scenario

In this fixture the authenticated north tenant requests beta, a document owned by south. The original lookup returns it. The corrected lookup does not. The missing tenant predicate—not the variable naming—is the defect.

Inspect the seeded defect and correction

Original fixture

documents.find(document => document.id === id)

Corrected core expression

documents.find(document => document.id === id && document.tenant === tenant)

Clean control: A north request for alpha must still succeed. A fix that blocks all documents is not acceptable.

An off-by-one claim needs the API contract

The stated API uses one-based pages. For [10, 20, 30, 40], page 1 should return [10, 20]; the original expression returns [30, 40]. With zero-based pages that original expression would not be the same bug.

Inspect the seeded defect and correction

Original fixture

items.slice(page * 2, (page + 1) * 2)

Corrected core expression

items.slice((page - 1) * 2, page * 2)

Clean control: The corrected version must return [30, 40] for page 2 and reject non-positive or fractional page numbers.

How product evidence would be collected

The same frozen changes, repository context and review instructions must be sent to each product. Record the product/model version, plan, configuration, date, latency and cost. Use multiple runs and a blinded assessment of unique, reproducible findings; deduplicate comments about the same root cause.

Report defects found and missed, confirmed findings divided by all substantive findings, false alarms on clean controls, and regressions introduced by proposed fixes. Publish denominators and uncertainty. A tiny, synthetic suite cannot establish performance on real production repositories.

Product study status: not run. No product rankings, precision/recall scores or test costs are claimed.

CodeRabbitGreptileCursor BugbotQodoGitHub Copilot Code ReviewGraphiteGitar
CodeRabbit
CodeRabbit · Essentials $24/user/mo billed annually ($30 monthly); Team $48 ($60 monthly); Advanced $72 ($90 monthly); Enterprise custom. Public open-source eligibility and plan conditions apply.
Autofix
Supported
Self-hosted
Enterprise only; not included in standard self-serve plans.
Merge gate
Supported
Greptile
Greptile · Starter: free, 1 active developer, 50 credits/month. Pro: $30/seat/month with 50 credits included per seat, then $1/extra credit. Free for qualified non-commercial open-source.
Autofix
Supported
Self-hosted
Enterprise offering; not the $30 Pro seat plan.
Merge gate
Supported
Cursor Bugbot
Cursor · Usage-based billing included with Cursor; 14-day free trial. The cited Bugbot page does not publish a per-review dollar rate.
Autofix
Supported
Self-hosted
Not listed
Merge gate
Not available
Qodo
Qodo (formerly CodiumAI) · Pro Team uses pooled credits at $0.012/credit (packs: 2,500 / 5,000 / 20,000 credits). 14-day trial with unlimited credits, no credit card required. Enterprise custom.
Autofix
Supported
Self-hosted
Vendor lists self-hosting; plan conditions need verification
Merge gate
Supported
GitHub Copilot Code Review
GitHub · Bundled with GitHub Copilot Pro ($10/mo) and Business/Enterprise ($19–$39/user/mo). Reviews for unlicensed PR authors bill via GitHub AI Credits plus Actions runner minutes.
Autofix
Supported
Self-hosted
Not listed
Merge gate
Not available
Graphite
Graphite · Starter $20/seat/mo and Team $40/seat/mo, both billed annually on current plan documentation (monthly billing raises the per-seat rate); Team includes unlimited AI reviews, merge queue and custom rules. Enterprise custom. Free Hobby tier for personal projects.
Autofix
Supported
Self-hosted
Unknown
Merge gate
Supported
Gitar
Gitar (Sonar) · Core $20/user/mo; Pro $40/user/mo (both capped at 50 users). Enterprise custom. Free Pro features for public OSI-licensed open-source projects.
Autofix
Supported
Self-hosted
Not listed
Merge gate
Supported

Documented profiles

Platform architecture & deployment notes

Vendor-documented configurations, billing conditions and limitations. No product defect-detection scores or first-hand usage claims are implied.

CodeRabbit

CodeRabbit

Vendor source ↗

Dedicated pull-request review with integrations across several Git hosts, plus CLI and IDE surfaces. Integration breadth is not evidence of defect-detection quality.

Deployment condition: Enterprise only; not included in standard self-serve plans. Plan source · checked 2026-09-04.

Pricing and rate-limit source checked 2026-09-04; other profile claims retain their earlier review date.

Config surface: .coderabbit.yaml at repository root and cloud dashboard
Review style: Agentic AI reviews on PRs and CLI with 1-click in-place fixes, learned repository patterns, linter/SAST integrations, and pre-merge checklist checks.
Hourly throughput: Rolling per-developer hour limit on the cited pricing page: Essentials 5/hr; Team 8/hr; Advanced 10/hr; Enterprise 12/hr.
Evidence limit: Precision, recall and false-positive burden on your repository are not established by feature or pricing documents.
Watch out: Current hourly limits are Essentials 5, Team 8, Advanced 10 and Enterprise 12 per developer. Legacy subscriptions can have different entitlements; the old Pro name should not be applied to new plans.
Hosts: GitHub, GitLab, Bitbucket, Azure DevOpsChecked

Greptile

Greptile

Vendor source ↗

Full-codebase architecture reviewer: builds an updated graph of the entire repo to catch ripple effects that standard diff-only linters miss.

Deployment condition: Enterprise offering; not the $30 Pro seat plan. Plan source · checked 2026-09-04.

Pricing and rate-limit source checked 2026-09-04; other profile claims retain their earlier review date.

Config surface: Custom repository instructions & rules in Greptile web console
Review style: Continuously indexes the full repository graph so PR reviews reason about call sites, type definitions, and dependencies outside the changed diff.
Hourly throughput: Usage is credit-metered rather than hourly. 1 standard review = 1 credit; 1 deep TREX cross-file review = 3 credits.
Evidence limit: Precision, recall and false-positive burden on your repository are not established by feature or pricing documents.
Watch out: Credits exhaust quickly if developers open frequent incremental PRs. GitHub and GitLab only. TREX reviews consume 3 credits each.
Hosts: GitHub, GitLabChecked

Cursor

Cursor Bugbot

Vendor source ↗

Cursor's native production reviewer: logic bug focused, directly connected to the developer's Cursor editor and background agents.

Config surface: .cursor/BUGBOT.md and team/repo rules
Review style: Bug-first automated code review. Prioritizes runtime panics, logic flaws, and race conditions over style nits. One-click sync fixes directly into the Cursor editor.
Hourly throughput: No hourly PR review cap. API endpoints rate-limited at 30–60 requests/minute. Reviews consume cloud agent usage credits.
Evidence limit: Precision, recall and false-positive burden on your repository are not established by feature or pricing documents.
Watch out: Default check is informational; blocking PR merges requires setting explicit fail-on-unresolved-issues policy. Git-host support beyond GitHub (GitHub Enterprise Server, GitLab incl. self-hosted, Bitbucket/Data Center, Azure DevOps Services) has deployment-specific conditions per the Bugbot docs; a supported self-hosted Git server is not the same as a self-hostable Bugbot service. Review billing is usage/credits (Autofix Cloud Agent credits are billed separately); the 30–60 req/min figure is an API rate limit, not a reviews-per-hour allowance.
Hosts: GitHub, GitHub Enterprise Server, GitLab, Bitbucket, Azure DevOpsChecked

Qodo (formerly CodiumAI)

Qodo

Vendor source ↗

Enterprise governance reviewer: combines automated PR reviews, test suite expansion, and cross-repo compliance policies.

Config surface: .pr_agent.toml, REVIEW.md, and Qodo cloud portal
Review style: Multi-agent review with cross-repo context, test generation, and shift-left agent skills that run both in the IDE and in CI pull request comments.
Hourly throughput: No hourly developer limit. Reviews draw from pooled workspace credits with configurable monthly budget caps.
Evidence limit: Precision, recall and false-positive burden on your repository are not established by feature or pricing documents.
Watch out: Credit packs require budget monitoring after trial. Do not confuse Qodo with Qodex (agentic QA platform).
Hosts: GitHub, GitLab, Bitbucket, Azure DevOpsChecked

GitHub

GitHub Copilot Code Review

Vendor source ↗

Zero-friction default for teams already paying for GitHub Copilot licenses: no third-party OAuth app or procurement needed.

Config surface: .github/copilot-instructions.md and AGENTS.md
Review style: Native GitHub PR integration. Can be requested or auto-assigned like a human reviewer. Proposes one-click suggestions directly in GitHub diff comments.
Hourly throughput: No hourly PR cap. Estimated at $0.05–$1 (Lite review) to $0.25–$5 (Balanced review) in AI credits plus Actions minutes for agent context gathering.
Evidence limit: Precision, recall and false-positive burden on your repository are not established by feature or pricing documents.
Watch out: Comments do not block merge natively. Drawing against pooled AI credits and Actions minutes can create billing surprises on large PRs.
Hosts: GitHub, Azure DevOps (preview)Checked

Graphite

Graphite

Vendor source ↗

Buy Graphite for high-velocity stacked diff workflows and merge queues; AI code review is included seamlessly in the Team tier.

Config surface: Graphite workspace automation rules and review filters
Review style: AI reviews tailored for stacked pull requests. Delivers fast summaries, diff checks, and proposed fixes inside Graphite's review inbox.
Hourly throughput: Hobby/Starter have capped review allowances; Team and Enterprise include unlimited AI reviews and automated CI summaries.
Evidence limit: Precision, recall and false-positive burden on your repository are not established by feature or pricing documents.
Watch out: GitHub only. Unlimited AI reviews require the Team plan ($40/seat/mo billed annually). Not designed as a standalone comment bot without adopting Graphite's PR workflow.
Hosts: GitHubChecked

Gitar (Sonar)

Gitar

Vendor source ↗

Autonomous CI healing reviewer: actively repairs failing tests and lints rather than merely leaving critique comments.

Config surface: .gitar/review/*.md, .gitar/rules/*.md, AGENTS.md, .cursorrules
Review style: Iterative reviewer: reviews the diff with call-site context, creates fixes, and executes CI in a loop until pipelines are green.
Hourly throughput: Manual '@gitar review' triggers are unlimited. Automatic background reviews and CI fix loops scale with assigned seats.
Evidence limit: Precision, recall and false-positive burden on your repository are not established by feature or pricing documents.
Watch out: Acquired by Sonar; confirm ongoing roadmap and SonarQube packaging before new enterprise commitments. GitHub Enterprise Server not supported.
Hosts: GitHub, GitLabChecked