AI code review tools in 2026: free, open source and paid

The leading AI code review tools in October 2026 are CodeRabbit, a dedicated reviewer from $24 a developer a month; GitHub Copilot code review, billed per review in AI credits; Claude Code, through Anthropic's managed Code Review or a free GitHub Action; Codex's @codex review; Cursor's Bugbot; and Greptile, Graphite, Qodo and Sourcery. For free and open source, PR-Agent is MIT-licensed and runs on GitHub, GitLab, Bitbucket and Azure DevOps with your own model key, and most paid tools are free on public repositories. Google's free Gemini Code Assist reviewer for GitHub shut down on July 17, 2026.

Key takeaways

• Billing is moving to usage: Copilot charges each review in AI credits, Bugbot moved from $40 a seat to about $1 to $1.50 a run, and Claude's managed reviews average $15 to $25 each.
• Open source: PR-Agent (MIT since July 2026) and Kodus run on any major Git host with your own model.
• Azure DevOps: CodeRabbit, Qodo, Bugbot, PR-Agent, Kodus, and Copilot in preview since August 26, 2026.
• Most AI reviewers comment rather than approve. Copilot can approve in preview, and CodeRabbit can approve once its comments are resolved.
• The benchmarks are mostly vendor-run. Martian's open benchmark is the closest to neutral, and no tool scored above 63% on its offline test.

Which AI code review tools are free?

• PR-Agent, the open-source reviewer Qodo started, now community-run and MIT-licensed since July 9, 2026. It works with GitHub, GitLab, Bitbucket, Azure DevOps and Gitea, and you bring your own model key (GitHub).
• Kodus, open source under AGPL-3.0 with a commercial edition, on the same hosts plus Forgejo.
• Claude Code's GitHub Action is free software, but each run spends GitHub Actions minutes and either API tokens or your Claude plan's usage (docs).
• Free on public repositories: CodeRabbit and Sourcery. Greptile is free for non-commercial open-source projects and has a free starter tier for one developer.
• Gone: Google's free Gemini Code Assist app for GitHub was deprecated on June 18 and shut down on July 17, 2026; the enterprise version continues (Google).

The paid tools

• CodeRabbit: $24 a developer a month billed annually ($30 monthly), $48 for Team and $72 for Advanced. It supports GitHub, GitLab, Bitbucket and Azure DevOps, and can approve a pull request once its comments are resolved if you turn that on (pricing). It raised $143 million at a $1.5 billion valuation in August 2026.
• GitHub Copilot code review: since June 1, 2026, each review costs AI credits, roughly $0.05 to $5 depending on effort, plus Actions minutes on private repositories. It can review pull requests from people without a Copilot seat on Business and Enterprise, and approving pull requests is in preview, off by default (GitHub).
• Claude Code Code Review: a research preview on Team and Enterprise. Reviews average $15 to $25, billed outside plan usage, take about 20 minutes, and never approve or block a pull request (docs). Locally, /code-review reviews your changes before you push.
• Codex: mention @codex review on a GitHub pull request, or turn on automatic reviews; it flags only the most serious issues and reads review rules from AGENTS.md. It draws on your ChatGPT plan's usage.
• Cursor Bugbot: usage-based since June 2026, at about $1 to $1.50 a run, on GitHub, GitLab, Bitbucket and Azure DevOps. It can be set to fail a check on unresolved issues. Cursor is now part of SpaceX, and it also owns Graphite.
• Greptile: $30 a seat with 50 review credits each, then $1 a credit, and it can be self-hosted. No Azure DevOps.
• Graphite: AI reviews in its pull request platform, $40 a user for unlimited reviews, GitHub only.
• Qodo: pooled credits, from $30 a month for about 18 reviews.
• Sourcery: $12 a developer a month, GitHub and GitLab.
• Atlassian Rovo Dev: $20 a developer a month for Bitbucket Cloud and GitHub; Atlassian is folding it into Jira.
• SonarQube adds AI checks to its quality gates, and bought the AI reviewer Gitar in May 2026.

How good are AI code reviewers?

Hard to say from public numbers. Most benchmarks are run by a vendor that ranks itself first: Greptile's own 2025 test put Greptile at 82% of planted bugs caught. The closest to neutral is Martian's Code Review Bench, open source since February 2026. On its offline set of 50 pull requests with 173 human-checked issues, no tool caught more than 63%; on real pull requests, CodeRabbit had the highest share of comments developers acted on (Martian).

The practical test is your own repository: run two tools on the same pull requests for two weeks and count the comments your team acts on.

AI review doesn't replace human review

Every vendor frames it the same way: the AI comments, a person decides. Anthropic's managed reviews never approve, Atlassian's tagline is "Reviewed by AI. Approved by you," and GitHub dismisses Copilot's approval when new commits arrive. As agents write more of the code, the human check matters more, and it works best before a change lands rather than after.

Poly (usepoly.co) puts that check inside the session: a team works with one coding agent in a shared room, the agent asks before it acts, and any member can approve or reject a change before it's made. How other tools handle approvals: Who approves a coding agent's changes? and Cross-model code review.