C

AI pull-request review · Independent review

From $24 per developer per month

CodeRabbit Review 2026: Benchmarks, Pricing and the Limits Nobody Advertises

Updated 31 August 2026 · Every figure re-checked against its source on 31 August 2026

Yes, if you want a second pass on every pull request and can tolerate waiting. Pro costs $24 per developer per month billed annually, $30 monthly. The default trial is Pro+ for fourteen days, no card required. Skip it if you need security scanning you can rely on alone, or if a review that sometimes lands 40 minutes later would break your workflow — one independent study put CodeRabbit’s mean at 9.5 minutes but its P95 at 41.3, the slowest of the four tools it tested.

What CodeRabbit actually is

CodeRabbit sits on your pull requests. It posts a summary, walks the diff line by line, runs a stack of linters that now includes zizmor, Vale, Verilator, Infer, React Doctor and SkillSpector, and and, on Pro+, a beta Fix CI feature that can push a commit for you: it launched on 16 July 2026 for GitHub, gained a fix-ci commit command on 3 August and Azure DevOps support on 19 August. Chat is a paid feature: Pro gives fifty chats, Pro+ and Enterprise a hundred. On the free tier a pull request gets a summary, not a full review — one an hour — though you still get three IDE and three CLI runs an hour, and the same 150-file cap. Worth knowing before you judge it on a free account., which is worth knowing before you judge it on a free account.

Since July 2026 it reaches beyond the pull request. The Security Agent, still beta, scans a whole repository on demand or on a schedule for vulnerabilities, leaked secrets and vulnerable dependencies, and covers GitHub, GitLab including self-hosted, Azure DevOps and Bitbucket Cloud. It is a separate add-on with its own fourteen-day trial; only AI Deep Scan is billed by usage. Triage, launched on 13 August 2026, is Pro+ and beta, limited to GitHub Cloud and GitLab Cloud.

Throughput is metered per hour and per file. Pro allows five PR reviews, five IDE and five CLI runs an hour, up to 150 files; Pro+ doubles the hourly allowance and lifts the cap to 300; Enterprise gets twelve an hour. Anything above 300 files is not reviewed at all, and you cannot buy your way past that with credits. Heavy weeks tighten things further: at sixty reviews in seven days, Pro drops to one an hour.

Three things it is not

It is not a replacement for SAST. On DeepSource’s OpenSSF CVE benchmark CodeRabbit scored 59.39% accuracy and an F1 of 36.19%. Resist the arithmetic: DeepSource defines accuracy differently from catch rate, the corpus mixes vulnerable and patched samples so true negatives count, and recall was never published, so no miss rate can be derived from those figures. The Security Agent may change the picture, but it is beta and sold separately.

It is not an architecture reviewer. In the dev.to study of 146 merged pull requests, CodeRabbit produced 281 findings across 82 of them, roughly 3.4 per PR on default settings. The study counted findings, false positives and latency rather than classifying their depth, but the shape of the product — comments anchored to a diff, plus hard file caps — points the same way: a review stops at 150 files on Pro, 300 on Pro+, and large refactors are exactly the changes where you most want a second opinion on structure. The file caps reinforce this. A review stops at 150 files on Pro, 300 on Pro+, and large refactors are exactly the changes where you most want a second opinion on structure.

It is not fast. The same study measured a 9.5 minute mean and a P95 of 41.3 minutes, the slowest of the four tools tested; Greptile averaged 4.9 minutes, Cursor BugBot 6.1, Seer 3.7 with a P95 of 6.8. One caveat worth stating: the author works at Sentry, and Seer posted the fastest times in the set. Even discounting that, a P95 of 41.3 minutes means some reviews land long after you have moved on.

The noise problem, and the settings aimed at it

Out of the box it talks a lot. The dev.to run recorded 281 findings across 82 pull requests on default configuration, and while the false positive rate was low at 2.3%, but volume is its own kind of cost: 281 comments is a lot to read past when the same corpus produced 120 from Greptile. Greptile returned 120 findings over the same corpus. More comments is not automatically better signal, and a reviewer nobody reads is a reviewer you are paying for and ignoring.

The fix is the quiet profile, shipped on 2 July 2026, which trims what the bot posts. Set it before you install, not after. A review bot has to earn its place early, and a stretch of noisy pull requests is the fastest way to lose that argument inside a team. Slop Detection, added for private repositories on the Pro plan on 21 August 2026, is worth turning on at the same time.

What the benchmarks actually say

Martian’s leaderboard does not ask whether a review comment was correct. Per Martian’s own description, it checks whether a developer edited code after receiving the comment. Precision is the share of a tool’s suggestions that match a real change; recall is the share of post-review changes the tool had flagged. That is a proxy for usefulness, grounded in developer behaviour rather than a judge model’s opinion — a different set of trade-offs, not automatically a better one, since an edit after a comment is not proof the comment caused it. It is not a measure of security coverage.

The definitions suggest where the bias would sit. Recall counts changes the tool mentioned, so a tool that comments on more lines has more chances to intersect with whatever the developer eventually edits — which should show up as a cost to precision rather than to recall. The board is consistent with that reading: of the three, CodeRabbit has the highest recall (58.0%) and the lowest precision (64.6%), Greptile the reverse (49.6% and 80.0%). It is a reading of the numbers, not something Martian states. If you are choosing between a chatty reviewer and a terse one, read recall as a proxy for coverage and noise together, and read precision as the column that penalises the chatter. Neither number tells you how many comments your team will ignore.

At launch, on January-February 2026 data, CodeRabbit’s own blog put it first on F1 at 51.2%, with precision 49.2% and recall 53.5%. Martian’s write-up backs the recall half independently, naming CodeRabbit as having the “highest recall in the online data” at 0.54. CodeRabbit also said its recall beat the next tool by almost 15%, without stating whether that means 15 percentage points or 15% relative. Treat the gap as unquantified.

On 30 July 2026 Greptile self-published a lead of its own: F1 60.8%, precision 76.2%, recall 50.6%. Read the live board on 31 August 2026, though, in its default ’Last month’ window with 17,717 PRs scored, and neither claimant is on top. Cubic Dev AI holds first place at F1 64.5%, precision 73.2%, recall 57.6%. Greptile sits sixth on 61.2%, CodeRabbit seventh on 61.1%. The launch headline no longer describes the board.

The ordering hides the shape. Of those three, CodeRabbit still has the highest recall at 58.0% against Greptile’s 49.6%, while Greptile has the highest precision at 80.0%. The ranking moves with the time filter and updates continuously, so any citation needs a date. One more trap: the roughly 300,000 PRs in CodeRabbit’s blog is its own review volume in that window, not the benchmark’s sample. Martian scored 5,035 CodeRabbit PRs. One figure counts traffic, the other counts graded reviews.

Read directly from the live leaderboard on 31 August 2026 with its default “Last month” window, across 17,717 scored pull requests. The board is a client-side app whose numbers change with the time filter and update continuously, so treat any ranking — including this one — as a reading with a timestamp, not a standing.

A 12 May 2026 dev.to study by a Sentry engineer ran four tools at default configuration over 146 merged PRs, producing 679 findings. CodeRabbit found the most: 281 findings across 82 PRs, about 3.4 per PR, with a false-positive rate of 2.3%. It was also the slowest of the four, averaging 9.5 minutes with a P95 of 41.3 minutes. Greptile returned 120 findings and no false positives across 118 judgements, in 4.9 minutes.

That is the trade-off as this one study measured it, at default configuration on 146 PRs: more findings and more waiting, against fewer findings returned faster and, on the 118 judgements reviewed, with no false positives. Cursor BugBot landed between them at 128 findings, 4.8% false positives and 6.1 minutes; Seer averaged 3.7 minutes with a P95 of 6.8. If your merge queue tolerates a wait, CodeRabbit’s volume is the argument for it. If reviewers are blocked on the check, that 41.3-minute tail is the argument against.

The company behind it

CodeRabbit announced a Series C of $143m on 12 August 2026 at a $1.5bn valuation, co-led by Atomico and Smash Capital. New investors include BMW i Ventures and Datadog. That follows a $60m Series B in September 2025 at a $550m valuation, bringing total funding to roughly $231m. For a buyer, the relevant point is runway: $231m raised and a $1.5bn valuation suggest the company is funded well beyond a typical contract term, though funding size is not a guarantee of continuity and says nothing about whether the reviews are good.

The traction figures come from the company itself, not an auditor. CodeRabbit says it has more than 17,000 customers, naming Adyen, Indeed, BMW, Nvidia, JFrog and Trivago, and claims more than two million reviews a week and revenue up more than fivefold year on year. Treat those as vendor claims with no independent verification; the named logos at least point to specific companies you can ask about in a reference call, while the growth multiple cannot be checked at all. If procurement needs audited numbers, ask for them before signing.

Harjot Gill, co-founder and chief executive, is now positioning the product as agentic change management rather than pull request review: Triage, Change Stack and CodeRabbit Security are the three pillars. Buyers should note what that means in practice. Triage is beta and limited to Pro+, Change Stack is early access and expected to land on Pro+, and Security is a separate add-on. If you are buying the pitch, note that none of the three pillars is generally available yet: Triage shipped on 13 August 2026 as a Pro+ beta limited to GitHub Cloud and GitLab Cloud, Change Stack is in early access and free during launch with Pro+ expected, and Security shipped on 9 July 2026 as a separate beta add-on with its own 14-day trial. They exist, but you are buying beta and early access, not settled, generally available features.

The 2025 remote code execution incident

In January 2025, after the 38C3 conference, researchers at Kudelski found that CodeRabbit could be made to run attacker-supplied code. The mechanism was mundane: a .rubocop.yml file can use require to load arbitrary Ruby, and CodeRabbit ran RuboCop outside its sandbox. Anyone able to open a pull request could therefore execute code on the vendor’s own infrastructure, a severe failure mode for a tool you point at every repository you own.

The blast radius was the environment: the exploit exposed Anthropic and OpenAI keys, a GitLab personal access token, Jira, Pinecone and PostgreSQL credentials, and the private key of the CodeRabbit GitHub App. That last item is the one that matters, because it carried write access to roughly one million repositories across more than 80,000 installations. A code review bot with write scope is a supply chain dependency, and this is what that dependency looked like when it failed.

The response was fast. Kudelski reported it on 24 January 2025; CodeRabbit disabled RuboCop within an hour and rotated the exposed secrets within three. The company says the permanent fix, running RuboCop inside the sandbox, reached production within twelve hours. One correction worth making, because it is widely misread: the 30 January date in Kudelski’s timeline is when CodeRabbit confirmed the fix back to the researchers, not the day the hole was closed.

Public disclosure came on 19 August 2025, seven months later, and CodeRabbit’s own account appeared the same week the Kudelski write-up spread. Hacker News commenters attacked that gap directly, and some said they had cancelled their subscriptions over it. Judge the vendor on both halves: the engineering response was measured in hours, while the public disclosure came seven months later, on 19 August 2025, in the same week the Kudelski write-up spread. Ask about disclosure policy before you sign. Ask about disclosure policy before you sign.

What the exploit reached: Anthropic and OpenAI keys, a GitLab personal access token, Jira, Pinecone and PostgreSQL credentials, and the GitHub App private key — which carried write access to roughly one million repositories across 80,000-plus installations.

1. January 2025 Kudelski Security finds the flaw after 38C3. A .rubocop.yml in a pull request uses require to load Ruby, and RuboCop runs outside the sandbox.
2. 24 January 2025 Reported. CodeRabbit disables RuboCop within an hour and rotates secrets within three.
3. Within 12 hours CodeRabbit says the permanent fix — RuboCop moved into the sandbox — reached production.
4. 30 January 2025 CodeRabbit re-confirms the fix to the researchers. This is a confirmation date, not the date the fix shipped.
5. 19 August 2025 Kudelski publishes. Seven months after remediation. CodeRabbit’s own account appears the same week.

What it costs, and the limits nobody advertises

CodeRabbit Pro is $24 per developer per month on an annual commitment, or $30 if you pay monthly. Pro+ doubles both figures: $48 annually, $60 monthly. The gap matters because several 2026 features sit behind Pro+ rather than Pro, so the annual figure you budget for is often the $48 one, not the $24 one. Lite and Pro Legacy stopped being sold on 8 June 2026, so Pro is the entry point for anyone buying today.

The trial defaults to fourteen days of Pro+ with no card required, which is what a new account gets unless it is provisioned otherwise. But trials also exist at Pro level, capped at five reviews an hour, and at Enterprise, capped at twelve, so do not assume every account is evaluating the same ceiling. Check which tier your trial actually runs at before you draw conclusions about throughput, because a Pro trial and an Enterprise trial will not behave alike.

Pro gives you five PR reviews, five IDE reviews and five CLI reviews per hour. Pro+ raises each to ten, Enterprise to twelve. The plans page lists PR, IDE and CLI limits as three separate numbers rather than one combined allowance, which reads as an advantage if your team splits work between the editor and CI. It stops helping the moment a busy afternoon needs more pull request reviews than the Pro cap allows in an hour; the documented limit is the ceiling itself, and the plans page does not spell out what happens to the requests beyond it, so run @coderabbitai rate limit to see where you stand rather than assuming they are simply held over.

File ceilings follow the same split: 150 files per review on Pro, 300 on Pro+ and Enterprise. Anything above 300 files is not reviewed at all, and no amount of extra credit buys past it. If your repository regularly produces large refactors, migrations or generated-code changes, that is a hard wall rather than a pricing decision, and on a monorepo it is one of the limits most likely to bite.

The least advertised constraint is fair usage. On Pro, your hourly cap slides down as weekly volume rises: five an hour up to 29 reviews in seven days, four at 30-39, three at 40-49, two at 50-59, and one an hour once you reach 60. Pro+ has its own table, bottoming out at one an hour at 90 and above. Run @coderabbitai rate limit in a pull request to see where you actually stand.

What the free plan really gives you

Signing up costs nothing and needs no card, though the plans page never actually prints a zero. The free plan is easy to misread. It allows one pull request an hour, but that request only gets a summary, not a full review. The free plan also gives three IDE reviews and three CLI reviews an hour, and it is the pull request path that is documented as summary-only; there is no chat with the bot at all. So if you are judging CodeRabbit’s review quality from a free GitHub account, you are judging the wrong thing entirely.

Open source is treated separately. Public repositories get between one and ten pull request reviews an hour depending on GitHub stars, so a well-known project is served far better than a new one. Below ten stars, reviews are not automatic at all and must be triggered by hand. For a young open source project, that means CodeRabbit is available but not ambient, which changes how much of your workflow you can build around it.

How it compares

On the Martian board read directly on 31 August 2026, with the default last-month window covering 17,717 scored PRs, Greptile holds the highest precision of the three at 80.0% against CodeRabbit’s 64.6%. CodeRabbit holds the higher recall, 58.0% against 49.6%. If your reviewers ignore noisy bots, Greptile’s precision matters more; if missed changes cost you, recall does. Both sit below Cubic Dev AI at the top.

Coverage is the sharper split. Greptile runs on GitHub and GitLab only, so Bitbucket or Azure DevOps teams can stop there. Its pricing is $30 per seat including 50 credits, with extra credits at $1 each, and a free Starter tier from 29 June 2026 that grants 50 credits a month, not 50 reviews: a trex review consumes three credits. CodeRabbit Pro is $24 per developer monthly billed annually, $30 billed monthly. Budgeting by seat differs from budgeting by credit; model your review volume before comparing the headline figures.

Cursor BugBot has dropped seat fees entirely and now bills by usage, roughly $1.00 to $1.50 per run, applied from the first renewal after 8 June 2026; Cursor Teams remains $40 per user monthly. It triggers when a pull request is updated rather than on every push, which caps spend somewhat. It also supports GitHub, GitLab, Bitbucket and Azure DevOps, the latter in limited availability. So platform breadth is not a CodeRabbit exclusive.

What CodeRabbit offers over Greptile is platform coverage, and over both is suggestion volume. In the May 2026 dev.to study of 146 merged pull requests on default settings, it produced 281 findings across 82 PRs, about 3.4 per PR, against 120 from Greptile and 128 from BugBot, with a 2.3% false positive rate. The cost is latency: 9.5 minutes on average and 41.3 minutes at P95, the slowest of the four tools measured. Greptile logged no false positives across 118 adjudications and returned in 4.9 minutes.

Who should choose it

It fits teams whose repositories sit outside GitHub, particularly on Bitbucket Cloud or Azure DevOps, and who want the Security Agent, which is in beta and covers GitHub, GitLab including self-hosted, Azure DevOps and Bitbucket Cloud as a separate add-on with its own 14-day trial. It also fits teams that want many concrete suggestions per pull request and can absorb a review that lands in minutes rather than seconds.

It fits badly if your changes are large or your volume is high. Reviews above 300 files are not supported at all, and paying with credits does not lift that. Pro throttles under fair usage: 30 to 39 reviews in seven days drops you to four an hour, 60 or more to one an hour. The free tier only summarises pull requests, with no full review and no chat.

Security-sensitive organisations should weigh the January 2025 incident. A .rubocop.yml in a pull request could load Ruby code, RuboCop ran outside a sandbox, and environment variables leaked, including the GitHub App private key that carried write access to roughly a million repositories. The response timeline, as the company describes it, was short: RuboCop disabled within an hour, secrets rotated in three, and a sandboxed permanent patch in production within twelve hours. Public disclosure came on 19 August 2025, seven months later.

Frequently asked questions

Is CodeRabbit worth it in 2026?

For teams on GitLab, Bitbucket or Azure DevOps it is one of the few mature options, and it finds more than its rivals. Pro is $24 per developer per month billed annually. Skip it if you need reviews inside five minutes, architectural feedback, or a security gate you can rely on by itself.

Is CodeRabbit still number one on the Martian benchmark?

No. It led on F1 when the benchmark launched with January to February 2026 data. Read on 31 August 2026, the live leaderboard put Cubic Dev AI first on F1 at 64.5%, Greptile sixth and CodeRabbit seventh. CodeRabbit still had the highest recall of the three at 58.0%.

Does the free plan review private repositories?

Not with a full review. The free plan covers public and private repositories but posts a pull-request summary rather than a review, one an hour. Full reviews on the free plan come through the IDE extension and the CLI, three an hour each, capped at 150 files.

What are the rate limits nobody advertises?

Fair usage tiers your hourly ceiling by how much you reviewed in the past seven days. On Pro, 0 to 29 reviews leaves you at five an hour; 60 or more drops you to one an hour. Reviews over 300 files are not supported at all, even if you pay for extra usage.

How do I stop CodeRabbit being so noisy?

Set the quiet profile, which shipped on 2 July 2026, before you install the app rather than after the first noisy week. CodeRabbit has not published before-and-after numbers for it, so treat it as the intended remedy rather than a measured one.

CodeRabbit or Greptile?

Greptile if you are on GitHub or GitLab and want fewer, higher-confidence comments: it led precision at 80.0% on the leaderboard reading of 31 August 2026, and returned no false positives across 118 judged findings in the DEV study. CodeRabbit if you need Bitbucket or Azure DevOps, which Greptile does not support, or if you would rather over-report than miss something.

Can CodeRabbit review pull requests written by coding agents?

That is now much of its pitch. Triage, released 13 August 2026 in beta for Pro+, ranks open pull requests across repositories, and Slop Detection flags pull requests that look like unreviewed AI output, reaching private repositories on Pro from 21 August 2026.

Verdict

CodeRabbit no longer leads the Martian benchmark. At launch it claimed the top F1 score of 51.2% on its own blog, with Martian independently confirming the recall figure; read live on 31 August 2026 it sits seventh, at 61.1% F1, behind Cubic Dev AI at 64.5% and Greptile at 61.2%. The board moves with the time filter and updates continuously, so treat any single ranking as a snapshot rather than a verdict.

Buy it for suggestion volume and for the platform coverage Greptile does not match, not for benchmark supremacy or the vendor’s own claims: revenue growth of more than five times year on year and “most installed AI App on GitHub” are self-reported, without audited figures. Note too that the dev.to study’s author works at Sentry, whose Seer was the fastest tool measured. A 14-day trial needs no card and defaults to Pro+, so run it on your slowest repository before committing.

Where every figure came from

Each number on this page was read off one of these pages on 31 August 2026. Two blind verification passes ran over the same list; where they disagreed, the figure was dropped rather than published.

• CodeRabbit plans, limits and fair-usage tiers · docs.coderabbit.ai
• CodeRabbit changelog · docs.coderabbit.ai
• CodeRabbit Security Agent documentation · docs.coderabbit.ai
• Martian Code Review Bench, live leaderboard · codereview.withmartian.com
• Martian, Code Review Bench v0 write-up · withmartian.com
• CodeRabbit on its launch-day benchmark result · coderabbit.ai
• Greptile on the July 2026 leaderboard reading · greptile.com
• Greptile changelog and pricing · greptile.com
• Four reviewers run in parallel, 146 pull requests · dev.to
• DeepSource on AI code review tools, OpenSSF CVE data · deepsource.com
• Kudelski Security, the CodeRabbit RCE write-up · kudelskisecurity.com
• CodeRabbit’s response to the disclosure · coderabbit.ai
• Cursor BugBot documentation · cursor.com
• Cursor on the BugBot billing change · cursor.com
• CodeRabbit Series C announcement · businesswire.com

Figures we did not publish

• A claim that CodeRabbit “misses roughly 41% of real vulnerabilities”. It is arithmetic on DeepSource’s accuracy figure, and accuracy is not recall. The miss rate would be 1 minus recall, which nobody publishes.
• AIMultiple scores of 1/5 completeness and 2/5 depth. Those numbers belong to a hypothetical “Tool B” inside an example block in AIMultiple’s published judging prompt, not to CodeRabbit.
• A count of user reviews next to our score. We have no user-review data, so there is no number to show.
• The names of two of the three co-founders. Only Harjot Gill, co-founder and CEO, is confirmed by sources we could open.

How this review was made

Desk research against primary sources, not a long-term hands-on deployment by this site. We read the vendor documentation, the live benchmark, the independent study and the security disclosure directly, and we ran two independent verification passes over every figure before publishing. We did not run CodeRabbit on a production codebase for this piece, and we do not claim otherwise.

Related on OfficeBooks

• Best AI code review tools, ranked and compared
• CodeRabbit vs Qodo, head to head
• The short CodeRabbit entry in our directory
• Every cloud and developer tool we have reviewed