8 AI Code Review Tools Worth a Trial, and What Each One Costs

For a 10-person team reviewing 150 pull requests a month, the seven paid tools below cost anywhere from about £140 to £2,800 at list prices. The most expensive one isn’t the one leading the independent benchmark, either: when I checked, that tool cost about £300.

So the price tells you how a tool bills, not how good it is. Some charge a flat seat, some charge per review, and some charge per run, which means every push that triggers a review costs again. That difference moves your bill far more than any leaderboard.

These are the eight I’d expect your engineers to bring you, roughly in order of how often they come up. I read the pricing pages, the docs, the security write-ups and the independent research for each one, and the prices are as of September 2026. Every vendor charges in dollars, so I’ve quoted dollars with pounds alongside, at roughly 75p to the dollar.

The eight at a glance

If your Git host rules some of them out, or your code can’t leave your own cloud, the filters narrow the list before you read any further.

1. CodeRabbit: the one your engineers have probably heard of

CodeRabbit is the most installed AI reviewer on GitHub’s Marketplace by a distance, with about 332,000 installs against 71,000 for the next one. It works on GitHub, GitLab, Bitbucket and Azure DevOps, and each pull request gets a summary, a walkthrough and inline comments, many of them with a one-click fix.

Essentials costs $24 (about £18) a developer on annual billing, or $30 (about £23) monthly, and you only pay for developers who open pull requests. Usage past the plan’s limits costs $0.25 a reviewed file, up to a cap you set. The free plan writes PR summaries but doesn’t review pull requests, which a lot of comparison articles still get wrong. Open-source projects get the Team plan free, but its privacy policy says it uses open-source code to train its systems, so the no-training promise covers your private code only.

A Sentry employee ran four reviewers side by side on 146 pull requests (Sentry sells one of them), and CodeRabbit found the most issues, with 2.3% false positives, and it was the slowest at 9.5 minutes a review. On Martian’s offline benchmark, though, only about a third of its comments hit a real issue, and the developer threads I read mostly complain about noise until it’s tuned. In 2025 Kudelski Security showed a pull request could run code on CodeRabbit’s own servers and read keys with write access to customer repositories. CodeRabbit fixed it within a week of the report and says it found no sign that customer data was accessed. Its servers are in the US, and I found no EU hosting option.

2. GitHub Copilot code review: the one you may already be paying for

If your team is on Copilot Business, you already have a reviewer. You add Copilot to a pull request like any other reviewer, or switch on automatic reviews for a repository. It only works on GitHub, with Azure Repos in preview, and GitHub says it now does one in five code reviews on GitHub.

Billing moved to usage in June 2026. Business is $19 (about £14) a user and includes $19 of usage, and GitHub’s own estimate is $0.25 to $5 a review at its Balanced effort level, or $0.05 to $1 on Lite. Reviews on private repositories also use Actions minutes. Copilot Free doesn’t include code review. On the individual plans, GitHub now trains on your interaction data unless you opt out, though Business and Enterprise are excluded, and there’s US or EU data residency.

On Martian’s online benchmark it was near the bottom across the full year and second in the month to 21 September, so the newer version looks much better, though a month is a short window. One setting to watch: since September, in preview and switched off by default, Copilot can approve pull requests and count towards required approvals. I’d leave that off.

3. Greptile: the easiest to start, the hardest to budget

Greptile indexes your whole repository into a graph of functions and calls, then reviews each pull request against it, so it can tell you a change breaks a caller three files away. You install the app, and the next pull request gets a summary, a confidence score and inline comments. It covers GitHub, GitLab and Bitbucket, but not Azure DevOps.

Starter is free for one developer with 50 credits a month. Pro is $30 (about £23) a developer, with 50 credits each and $1 for every extra credit. A Base review costs 1 credit, Plus costs 3 and Apex costs 10, and Greptile recommends Apex on anything merging into main. So ten developers doing 150 Base reviews a month pay $300, all inside the allowance, and the same month on Apex is about $1,300 (roughly £980). Credits aren’t pooled, so one busy developer can pay overage while the rest of the team is under. Pre-Series A startups with under $2M revenue get 50% off.

The quality evidence is good. The Sentry employee’s trial marked none of its 120 findings as not applicable, and it had the highest precision of any tool in Martian’s latest online month. Two things to check before you sign: it keeps your code on its servers until you revoke access, and its security page says it may use anonymised customer data to train models unless you opt out.

4. cubic: the newcomer at the top of the benchmark

cubic is a small London startup from Y Combinator’s 2025 intake, and it’s the tool that led Martian’s online benchmark in the last week, the last month and the last 60 days when I checked. It came second on the offline test too. It only works on GitHub.

Starter is free with 20 pull request reviews a month, and open-source projects are free outright. Team is $30 a developer on annual billing, or $40 (about £30) monthly. Its AI providers are contractually barred from training on your code.

The catch is that it’s young. It has SOC 2 Type I rather than the Type II most bigger vendors hold, and almost nobody talks about it outside its own marketing. So I’d treat it as promising rather than proven, and it’s the tool I’d put up against whatever else you’re trialling.

5. Cursor Bugbot: the one for teams already living in Cursor

Bugbot needs a Cursor plan, from $20 (about £15) for an individual or $40 (about £30) a seat on Teams. Since June 2026 the reviews themselves are billed by usage, at about $1 to $1.50 a run. It runs on GitHub, GitLab (including self-managed), Bitbucket Data Center, and Azure DevOps and Bitbucket Cloud in beta.

The word that matters there is run. Each push that triggers a review is billed again, and one user on Cursor’s forum reported 7 to 10 runs per pull request. If your team pushes small commits often, set it to review once or on request.

SpaceX completed its purchase of Cursor in August 2026, which is worth a fresh look at the data terms if you’re a regulated business. Privacy Mode means no training on your code. And in August Lumia Security showed that running Cursor’s local review command on a hostile fork’s branch can run the attacker’s code on the reviewer’s machine. Cursor treated it as out of scope. The cloud reviewer isn’t affected, but review untrusted forks from the diff rather than by checking them out.

6. Claude Code Review: thorough, and priced like it

Anthropic’s managed reviewer is still a research preview, for Team and Enterprise plans on GitHub only. Several agents review each pull request, check each other’s findings and post the result as a neutral check, so it never blocks a merge. A review takes about 20 minutes.

It averages $15 to $25 a review (about £11 to £19), billed as usage on top of the plan. At 150 pull requests a month that’s £1,700 to £2,800, and more if every push re-triggers a review. There’s a monthly spend cap, and I’d set it before the first pull request.

The cheaper route is Anthropic’s GitHub Action, which runs Claude in your own GitHub Actions and bills only the model tokens. It can use Bedrock or Vertex in your own region too. In May 2026 it needed a security fix for a flaw where a malicious pull request could run code on the runner. It was fixed in 1.0.74, and anyone on the @v1 tag got the fix automatically. Both routes post as the same bot on GitHub, so Martian can’t separate them, but together they were fourth in the latest online month.

7. Qodo: the one built for bigger companies

Qodo is a multi-agent reviewer with a judge that filters the findings, a central rules system and context from across your repositories. It covers the most Git hosts of any tool here: GitHub, GitLab, Bitbucket, Azure DevOps and, on Enterprise, Gerrit. Its higher-effort Extended mode came top of Martian’s offline test.

It doesn’t have a free tier any more, only a 14-day trial. After that you buy credit packs, from $30 (about £23) for about 18 reviews up to $240 for about 144, for teams of up to 30. The free route is PR-Agent, the open-source project Qodo handed to the community in April 2026. It’s MIT licensed, works with any model including a local one, and is a simpler single-pass reviewer.

For a larger or regulated business, its paperwork is the strongest here. Its data processing addendum offers EU hosting and a UK variant of the standard contractual clauses, it says it doesn’t train on your code, and Enterprise adds single-tenant, on-premises and air-gapped options. Its security history is worth a question in procurement, though. In late 2024 Kudelski Security showed prompt injection could leak a GitHub token, and a later write-up showed Qodo’s own AWS key, with administrator access, leaking twice. Qodo says the key was for a development environment with no customer data, and everything was fixed and rotated by October 2025.

8. Alibaba’s Open Code Review: free, and yours to secure

Alibaba open-sourced its internal reviewer in May 2026. It’s a single command-line tool under the Apache-2.0 licence that runs in your own CI on GitHub, GitLab, Bitbucket or Gerrit, and you point it at whichever model you like, including Bedrock in your own region or a local model. You pay the model bill and your CI minutes, which on Alibaba’s own figures is 330,000 to 740,000 tokens a review.

It’s built for precision over recall. On Alibaba’s own benchmark it found 20% of the known issues, and about a third of its comments were on real ones. The one outside test I found, on 10 pull requests, saw far more noise than that, and nobody has re-run it since the maintainers fixed the bug behind it.

Security is where you pay for free. The npm version updates itself by default, and it had 134 releases in its first four months. In September it fixed a flaw, disclosed properly in an advisory, that let a malicious repository read files from the reviewing machine. It doesn’t redact secrets before sending code to the model, and several built-in model presets point at endpoints in mainland China. If you run it, pin the version, set OCR_NO_UPDATE=1 and restrict which providers your team can pick.

What your team would pay

The worked examples above use a 10-person team, so put in your own numbers here. The figures are list prices, and the notes say what each one leaves out.

Also worth knowing

• If you’re on GitLab, GitLab Duo’s reviewer costs a flat $0.25 (about 19p) a review, on GitLab.com and self-managed. That’s the simplest answer for a GitLab team.
• If you already pay for Sentry, its Seer reviewer uses your production errors to predict which changes will break. It’s $40 per active contributor a month on top of a paid Sentry plan, and it deletes your code after each run.
• Graphite agreed to be bought by Cursor in December 2025. It’s a good tool for stacked pull requests, but its reviewer came last on Martian’s offline test, catching under 8% of the known issues.
• Some old favourites have gone. Google shut down the free Gemini Code Assist app for GitHub in July 2026, Amazon’s CodeGuru Reviewer stopped taking new repositories in November 2025, and Boost Security bought Korbit in May 2026. If a roundup still recommends them, it hasn’t been updated.

Every vendor is number one

I’d put very little weight on any vendor’s chart. Martian’s benchmark is independent, and CodeRabbit, Qodo, cubic and Greptile have each quoted a different window or mode of it to claim first place. When I checked, cubic led the last month and Greptile led July.

The independent research is sobering. One recent study found the best model scored an F1 of 0.847 on synthetic bugs and 0.066 on real pull requests. Across the models tested, scores collapsed on big diffs too, from 0.657 under 10 lines to 0.043 over 150. At Mozilla and Ubisoft, reviewers accepted about 8% of AI comments. And at Beko, a PR-Agent-based reviewer got most of its comments resolved, but the average time to close a pull request went from under six hours to over eight.

So the only benchmark that counts is your own pull requests, and small pull requests help the bot as much as they help your team.

The round table

To pressure-test a recommendation, I ran the research past three archetypes. There’s a CTO at a 25-person Series A company, a QA lead at a company the same size, and an infosec lead at a regulated UK fintech. They’re composites, not real people, but each one holds the priorities I hear from those roles.

They agreed on more than I expected. None of these tools should be a merge gate: they leave advisory comments and a human approves. The bot gets read and comment permissions only, it never approves or merges, and no workflow a fork can trigger gets secrets. That’s because CodeRabbit, Qodo, Anthropic’s GitHub Action and other AI review and coding tools, including Cursor’s local review command, have all had code-execution or prompt-injection findings published against them. And the trial has to be on your own code: in that Sentry trial, 93.4% of the places flagged were flagged by only one of the four tools.

They disagreed in three places.

• SaaS or self-run. The CTO wants SaaS, because a reviewer nobody has to maintain beats a better one somebody has to babysit. The infosec lead wants self-run or single-tenant, with the model in the UK or EU and no training on code. The QA lead doesn’t mind, as long as the comments are worth reading.
• Precision against recall. The QA lead thinks a tool that finds 20% of known issues is too thin to count as a safety net. The CTO would take that trade if the comments are right, because low noise is what keeps developers reading them. The 8% acceptance rate at Mozilla shows what noise does.
• Per-review billing. The QA lead wants the deepest review on anything going into main, because that’s where a miss costs most. The CTO doesn’t like a bill that scales with every push. They landed on a spend cap, the cheapest tier by default, the deepest tier only for release branches or high-risk paths, and a review when the pull request opens or on request rather than on every push.

What I’d pick

Some larger businesses have a platform team and a firm rule against sending code to any new vendor. For them, Alibaba’s tool or Anthropic’s GitHub Action, running against Bedrock in the UK or EU region of the cloud account they already have, is a real option. It needs the same hardening as above, plus someone who owns it.

Before you roll any of them out

Run two tools on the same pull requests for two weeks, keeping the pull requests small while you do it. Count how many comments led to a code change and how many got ignored, because that acceptance rate tells you more than any leaderboard and it’s about your code. Keep the tool whose comments your team acted on more often, and if neither gets well above the 8% that Mozilla and Ubisoft saw, don’t buy either yet.

Then decide who owns the reviewer. It’s a vendor with access to your whole codebase and a comment box on every change. Somebody should be answerable for its permissions, its bill, and whether anyone’s still reading its comments three months in. I wrote about that gap in the review layer post: the tool helps, but it doesn’t replace the person who can say no.

If you’d like a second pair of eyes on the trial or the rollout, book a call.