AI Coding Tools
Which AI code review tool for a GitHub team, and what it costs
On this page
Anthropic took Claude Code Review out of research preview on October 15, so a GitHub team now has three serious AI reviewers to choose from: Claude Code Review, GitHub Copilot code review, and Cursor Bugbot. All three bill per review now. None of them publishes a fixed price per pull request.
Which AI code review tool fits a GitHub team comes down to three things: what your team already pays for, where your code is hosted, and how often reviews run. The last one matters most for cost, and it is the one most teams leave at the default. On two of the three tools, the default is “review every push.”
The default setting decides the bill
Here is how each tool runs out of the box, from each vendor’s own docs:
| Default trigger | Default depth | Who pays | |
|---|---|---|---|
| Claude Code Review | After every push | Standard (of Quick, Standard, Deep) | The organization, through usage credits |
| Copilot code review | Once per PR, unless “Review new pushes” is on | Balanced (of Lite, Balanced) since Sept 28 | The PR author’s credits by default, or the org if an owner chooses |
| Cursor Bugbot | Every push | Default (of Low, Default, High, Smart) | Plan usage, then on-demand spend |
Anthropic’s cost page is blunt about it: a repository added from the admin page “starts at After every push, which runs the most reviews,” and that setting ends up “multiplying cost by the number of pushes.” A PR that gets pushed four times before merge is up to four reviews, not one.
Copilot is the odd one out. GitHub’s docs say that unless it has been configured to review each push, Copilot “will only review a pull request once.” What changed on the GitHub side is depth: on September 28, the Default effort level moved from Lite to Balanced for every repository that had not explicitly picked Lite. GitHub says Balanced reviews use more AI credits.
What each reviewer does, and where it runs
Claude Code Review sends several agents through the diff and the surrounding code in parallel, runs a verification step to drop false positives, then posts inline comments, a summary, and a check run. It runs on Anthropic’s infrastructure, reads the repo read-only, and does not build or run your code. It focuses on bugs that would break production rather than style, and you steer it with a REVIEW.md file that it reads from the base branch, never from the PR itself. So a contributor cannot rewrite the review rules in the same PR they want reviewed.
The limits are real. It needs a Claude Team or Enterprise plan and GitHub (Enterprise Server works). It is not available to organizations with Zero Data Retention or the HIPAA configuration, it is not covered by Anthropic’s BAA, and it does not yet support customer-managed encryption keys. You cannot pick the model either; the review level is the only dial.
Copilot code review draws from the same AI credit pool as the rest of Copilot, plus GitHub Actions minutes for its context gathering (self-hosted runners do not burn minutes). It can be requested through the REST and GraphQL APIs since October 2, with an effort level per request. Business and Enterprise orgs can also turn it on for members without a Copilot license, billed straight to the org.
Cursor Bugbot works on GitHub, GitLab, Bitbucket and Azure DevOps, with Bitbucket Cloud and Azure DevOps still in beta. It can hand findings to a Cursor agent to fix (Autofix), which bills separately as Cloud Agent usage. It switched from a $40 seat to per-run billing in May 2026.
One team, three bills
Take a team of eight developers opening 120 pull requests a month, with an average of three pushes per PR before merge. These are the per-review figures each vendor has put out:
- Claude Code Review: $15 to $25 per review, from Anthropic’s March 2026 launch post. The GA docs dropped the number and say “no flat per-review price.”
- Copilot code review: $0.25 to $5 of AI credits at Balanced, $0.05 to $1 at Lite, from GitHub’s docs, excluding Actions minutes. One credit is $0.01.
- Cursor Bugbot: $1.00 to $1.50 per run, Cursor’s stated average as reported by a competitor’s review of its docs. Cursor’s pricing page lists no rate.
Run those through the same month:
| Setup | Reviews per month | Estimated review cost |
|---|---|---|
| Claude Code Review, default (every push) | up to 360 | $5,400 to $9,000 |
| Claude Code Review, once per PR | 120 | $1,800 to $3,000 |
| Copilot code review, default (once per PR, Balanced) | 120 | $30 to $600, plus Actions minutes |
| Copilot code review, once per PR, Lite | 120 | $6 to $120, plus Actions minutes |
| Bugbot, default (every push) | 360 | $360 to $540 |
| Bugbot, once per PR | 120 | $120 to $180 |
These are review costs only. They sit on top of seats: Claude Team is $20 per seat per month billed annually ($25 monthly), Copilot Business is $19 per user per month, and Bugbot needs a Cursor plan. Claude Code Review usage is billed outside the Team plan’s included usage, so the seats do not absorb it.
Two caveats keep the table honest. Claude’s every-push number is a ceiling, because Anthropic’s docs say some pushes keep the earlier result instead of running a fresh review. And every range moves with PR size: Anthropic’s own internal data shows PRs over 1,000 lines get findings 84% of the time against 31% for PRs under 50 lines, and bigger diffs cost more on every tool.
The point is not the exact dollars. It is that the trigger setting swings each bill by roughly 3x, and the vendor choice swings it by an order of magnitude.
Which AI code review tool fits your team
You already pay for Copilot Business or Enterprise. Turn on Copilot code review for every repository first. At 120 PRs a month the review spend is in the tens to low hundreds of dollars, and an eight-seat Business org pools 15,200 included credits a month, worth $152, though chat and agent usage draw from the same pool. Watch that pool after September’s switch to Balanced before you buy anything else.
You have a few repositories where a missed bug costs real money. Payments, auth, the service everything else calls. That is what Claude Code Review is priced for. Anthropic’s launch post says that internally the share of PRs getting substantive review comments went from 16% to 54%, and that engineers marked under 1% of findings incorrect. Those are vendor numbers from its own codebase, but the design (multiple agents plus a verification pass, about 20 minutes per review) is aimed at depth, not volume. Turn it on for those repos only, set them to once per PR, and use Deep where the risk justifies it.
Your code is on GitLab, Bitbucket or Azure DevOps. Claude Code Review and Copilot code review are out. Bugbot is the managed option of the three. If you would rather keep Claude, Anthropic documents running Claude in your own GitLab CI/CD with an API key, where you pay API token rates and own the pipeline.
You are regulated, on ZDR, or need CMEK. Claude Code Review is not available to you today. Copilot Business and Enterprise, or Claude in your own CI on your own runners, are the paths that remain.
You are a two-person team on Claude Pro or Max. The managed service is not available on those plans. Run /code-review in a Claude Code session before you push; Anthropic lists it as available on any plan, and it draws from your plan’s normal usage instead of a separate bill. If you are routing that kind of work to cheaper models, our Haiku 5.5 subagent guide covers the setup.
Set it up so the first invoice is not a surprise
Whichever tool you pick, do these four things before you add your whole org.
Set a spend cap first. On Claude, the Code Review limit lives at claude.ai/admin-settings/usage, and Anthropic will not let you cover a whole GitHub organization until a monthly limit exists. When the budget runs low, reviews are skipped and the PR gets a comment saying why. On GitHub, set a budget for AI credits paid usage. On Cursor, cap on-demand spend.
Change the trigger before the first PR, not after the first bill. Claude: Admin settings > Claude Code > Code review, Edit on each repository row, and set Review runs to “Once after PR creation.” Bugbot can be set per repo in a config file it reads from the base branch:
# .cursor/config/bugbot.yaml
version: 1
triggers:
frequency: oncePerPr # default is everyPush
review:
effort: default # low | default | high | smart
On Claude, developers can still opt one PR into per-push reviews by commenting @claude review always. That is the right escape hatch for a risky PR, and it is also how the bill creeps back up.
Pilot, then multiply. Anthropic’s own advice is the right method for all three: enable a few representative repositories, let a couple of weeks of reviews run, read the average cost per review (Claude shows it in the Repositories table), and multiply by your monthly PR count, then by pushes per PR if you stay on every push.
Write the review rules down. REVIEW.md for Claude, .cursor/BUGBOT.md for Bugbot, and repository custom instructions for Copilot. GitHub warns that custom instructions raise credit consumption, so keep them short. A reviewer that knows your conventions flags fewer false positives, and false positives are what make a team stop reading the comments. If your team keeps rewriting the same review prompts by hand, the AI Prompt Library has code review templates you can adapt into these files.
None of the three blocks a merge by default. Claude’s check run never concludes as a failure, and Bugbot’s concludes neutral. Copilot is the only one that can approve a PR, and only when an admin turns Copilot approvals on. Keep a human approval rule in branch protection regardless.
Bottom line
For most GitHub teams the answer is Copilot code review on everything, because it reviews once per PR by default and costs cents to a few dollars a review. Add Claude Code Review only on the handful of repositories where a production bug is expensive, set those to once per PR, and cap the spend before you start. Pick Bugbot if your code is not on GitHub or your team already lives in Cursor, and switch its trigger off every push on day one.
The vendor you pick matters less than the default you forget to change. Check the trigger on every repository today, and compare what you find with the Copilot plan breakdown before the next billing cycle.
Frequently Asked Questions
How much does Claude Code Review cost per month?
There is no flat price. Anthropic bills each review on token usage through usage credits, separate from your Team or Enterprise seats. Its March 2026 launch post put the average at $15 to $25 per review; the current docs give no figure and tell you to pilot a few repositories and read the average cost per review from the admin page.
Is Copilot code review cheaper than Claude Code Review?
Per review, by a wide margin on GitHub’s own estimates: $0.25 to $5 of AI credits at the default Balanced effort and $0.05 to $1 at Lite, plus GitHub Actions minutes. Claude Code Review is built as a deeper, slower pass, so the fair comparison is Copilot on every PR and Claude on the repositories where a missed bug is expensive.
Does Claude Code Review work with GitLab or Bitbucket?
No. It supports GitHub only, including GitHub Enterprise Server. On GitLab you can run Claude in your own GitLab CI/CD with an API key, or use Cursor Bugbot, which supports GitHub, GitLab, Bitbucket and Azure DevOps, some of them in beta.
Can an AI code reviewer block a merge?
Not by default on any of the three. Claude Code Review never approves or blocks, and its check run does not conclude as a failure. Bugbot’s check concludes neutral unless you turn on failing for unresolved issues. Copilot can approve when an admin enables Copilot approvals, but it does not block.
Share
Grant M.
Developer with IBM i and full-stack experience. Covers AI tools and automation for software developers at PromptedDev, with a focus on real workflows, honest comparisons, and legacy system modernization.
Related reading