Veridical Updated By Łukasz Szafrański

AI code review tools, compared on the same pull requests

We compared Veridical with 5 AI code reviewers on real merged pull requests they had already reviewed, and we print every result with its lens, interval and date. On the credited-BT lens Veridical scored higher than CodeRabbit, cubic and Greptile; on the strict lens CodeRabbit, Gemini Code Assist and Greptile scored higher.

Pick by Git host first and by what backs each comment second: every competitor fact below links to its source.

Comparisons

Measured on the same pull requests

Credited-BT first, strict beside it. Each row is its own set of shared pull requests, so rows are not comparable with each other. We claim a difference only when the 95% interval excludes zero and at least 30 pull requests are shared. OpenAI Codex code review (8 pull requests) did not reach that minimum, so its row is printed but not read. Measured on credited-BT, not claimed: F1 67.5 for Veridical and 58.6 for Gemini Code Assist (95% CI plus 0.8 to plus 18.0). We print it without claiming a difference: the interval starts less than one point above zero and our claim rule lets us claim a lead only over CodeRabbit, cubic and Greptile.

ToolShared PRsCredited-BT F1, Veridical vs toolStrict F1, Veridical vs toolCredited-BT resultStrict result
CodeRabbit5670.6 vs 57.736.1 vs 52.2Veridical ahead by 12.9 points (95% CI 4.9 to 21.6)CodeRabbit ahead by 16.1 points (95% CI 5.5 to 27.2)
cubic4371.9 vs 50.838.2 vs 40.1Veridical ahead by 21.0 points (95% CI 13.5 to 29.3)No difference claimed (95% CI minus 11.7 to plus 8.3)
Gemini Code Assist3267.5 vs 58.638.2 vs 56.9Measured, not claimed (95% CI plus 0.8 to plus 18.0)Gemini Code Assist ahead by 18.7 points (95% CI 5.3 to 30.7)
Greptile6470.6 vs 58.437.2 vs 49.1Veridical ahead by 12.2 points (95% CI 6.9 to 17.9)Greptile ahead by 11.9 points (95% CI 3.7 to 19.5)
OpenAI Codex code review874.7 vs 59.542.0 vs 51.1Too few pull requests to read (95% CI plus 3.0 to plus 31.6)Too few pull requests to read (95% CI minus 24.4 to plus 4.8)

Run coverage differs by comparison and by side. No run was attempted on Greptile's unmatched findings, or on either side's in the OpenAI Codex code review comparison, so those were scored by the grader alone. The counts are on the false-positive page.

Credited-BT
A comment counts as correct when it matches a change the author made after review, or when a grader judged the issue real. Where the issue could be built and run, the run overrides the grader: confirmed counts, refuted does not. The same rules apply to both sides, but how many findings a run was attempted on differs by comparison and by side, so each comparison prints its own counts.
Strict
A comment counts as correct only when it matches a change the author made after review. The tool's own comments were visible to the author and ours were not, so this lens favours the tool whose comments the author saw.

How this was measured

  • Readout readout_20260720_191140.json, generated , protocol 1.6.2, scored by offlinebeat/h2h/score_h2h.py. Precision and recall are averaged per pull request; F1 is computed from those averages. The 95% confidence interval (CI) is a paired bootstrap over pull requests with 3000 resamples.
  • The study was pre-registered on with 3 comparators (CodeRabbit, cubic, Gemini Code Assist) and later extended to 5 (Greptile, OpenAI Codex code review added). The pre-registered headline was the strict lens. We lead with credited-BT because strict scores comments against changes the author made after seeing the other tool's comments.
  • Pull requests: merged since March 1, 2026 in public repositories where the tool posted at least one comment, with 2 to 12 commits, at most 15 changed files and fewer than 800 added lines. Veridical reviewed the same pull requests afterwards, from the state before review.
  • Each comparison is still below the protocol's first planned checkpoint of 150 pull requests. These readings come before that checkpoint, so treat them as early.
  • The Veridical side was our July 2026 research pipeline. It predates the production engine that reviews pull requests today, so read these as dated measurements, not as a statement about today's reviews.
  • The CI covers which pull requests were sampled. It does not cover grader variation: a model grades the findings no author acted on, and grading the same findings again can move credited scores. We have not yet published a replicate-grading measurement, so treat small credited differences with care.
  • We claim a difference only when the 95% CI on the named lens excludes zero and at least 30 pull requests are shared. A lead for Veridical is claimed only over CodeRabbit, cubic and Greptile, the tools our claim rule names, and only when its interval starts at least one point above zero; any other lead for Veridical is printed as measured, not claimed. A lead for the other tool is stated whenever its interval excludes zero. Otherwise we print the numbers and claim neither side.
  • A difference is Veridical's F1 minus the tool's, computed before rounding, so it can differ in the last digit from the gap between the two printed F1 values. A claimed difference is printed with its interval oriented to the side that scored higher; every other interval carries a plus or minus sign.

Snapshot

Tools in A to Z order. Order is not a ranking. Prices and terms come from each vendor's own pages, checked on ; follow the links for current details. The measured column comes from our same-PR benchmark and stays empty for tools we did not measure.

ToolGit hostsPricingFree or open-source termsWhat backs a findingMeasured result
CodeRabbitGitHub, GitLab, Azure DevOps, Bitbucket Cloud and Bitbucket Data Center (docs.coderabbit.ai)Per developer: Essentials $24, Team $48, Advanced $72 a month billed yearly (Essentials $30 and Team $60 billed monthly). Enterprise on request. Only developers who open pull requests are charged. (coderabbit.ai, checked )Free reviews on public repositories (“free reviews forever for public repositories”). 14-day trial on paid plans. (coderabbit.ai, checked )The model's review, plus output from 50+ linters and security analysers that “run automatically in secure sandboxed environments”. (docs.coderabbit.ai)Credited-BT: Veridical ahead by 12.9 points. Strict: CodeRabbit ahead by 16.1 points. 56 shared pull requests, readout 2026-07-20.
cubicGitHub only (“We currently only support GitHub”) (docs.cubic.dev)Per developer: Team $40, Pro $99, Max $200 a month billed monthly, with a discount for yearly billing. Each plan includes a monthly allowance of reviewed lines per developer. (cubic.dev, checked )Free for public repositories (“cubic is free for public repositories”), a free Starter tier, and a 7-day trial with no card. (cubic.dev, checked )The model's review against the whole codebase and the team's rules (“Every PR is checked against your whole codebase and your team’s rules”), plus a merge confidence score from 1 to 5. (cubic.dev)Credited-BT: Veridical ahead by 21.0 points. Strict: No difference claimed. 43 shared pull requests, readout 2026-07-20.
Gemini Code AssistGitHub, including GitHub Enterprise Server and Enterprise Cloud on the enterprise version (docs.cloud.google.com)Google's docs say “there are no charges for using Gemini Code Assist on GitHub during Preview”. A Google Cloud project with billing is required. (docs.cloud.google.com, checked )No charge during the Preview, per Google's docs. (docs.cloud.google.com, checked )The model's review, using information it “will automatically retrieve” from the repository and pull request, and an optional style guide. (docs.cloud.google.com)Credited-BT: Measured, not claimed (95% CI plus 0.8 to plus 18.0). Strict: Gemini Code Assist ahead by 18.7 points. 32 shared pull requests, readout 2026-07-20.
GitHub Copilot code reviewGitHub.com, with Azure DevOps in public preview (docs.github.com)Included from Copilot Pro at $10 per user a month (Pro+ $39, Max $100). Reviews consume GitHub AI Credits, and agentic reviews also use GitHub Actions minutes. (github.com, checked )Pull request reviews are not included in Copilot Free. (github.com, checked )The model's review with full project context gathering run on GitHub Actions. GitHub's own advice: “Always validate Copilot's feedback carefully.” (docs.github.com)Not measured by us.
GreptileGitHub, GitLab and Bitbucket (cloud and self-managed), plus Gitea; self-hosting is offered (greptile.com)Starter is free for one developer. Pro is $1 per credit with no seat fees: a base review is 1 credit, Plus 3, Apex 10. Enterprise on request. (greptile.com, checked )Free for qualifying non-commercial open-source projects with an OSI-approved licence and at least 50 stars, by application. 14-day Pro trial. (greptile.com, checked )The model's review with codebase context. TREX, in beta, starts the services, exercises the changed code and “leaves the logs, screenshots, and traces on the PR”. (greptile.com)Credited-BT: Veridical ahead by 12.2 points. Strict: Greptile ahead by 11.9 points. 64 shared pull requests, readout 2026-07-20.
OpenAI Codex code reviewGitHub (learn.chatgpt.com)Part of ChatGPT plans: Plus at $20 a month lists “Cloud-based integrations like automatic code review”; Pro from $100 a month. (learn.chatgpt.com, checked )Automatic code review is first listed on the Plus plan. (learn.chatgpt.com, checked )The model's review of the diff, following AGENTS.md guidance. On GitHub it “flags only P0 and P1 issues”. (learn.chatgpt.com)Credited-BT: Too few pull requests to read. Strict: Too few pull requests to read. 8 shared pull requests, readout 2026-07-20.
QodoGitHub, GitLab, Bitbucket and Azure DevOps, plus Gerrit (docs.qodo.ai)Credit packs pooled across the team, from $30 a month on Pro Team; Enterprise on request. 14-day trial with no card. (qodo.ai, checked )Free code review for actively maintained public GitHub repositories with 200+ stars, or in an organisation with a public repository that has 200+ stars. To start, install the Qodo app from GitHub Marketplace. (docs.qodo.ai, checked )Multi-agent review against organisation rules, with cross-repository review (“Catch breaking changes before they reach other repos”). (qodo.ai)Not measured by us.
VeridicalGitHub (github.com) only. GitLab and Bitbucket are not supported yet. (veridical.dev)Standard $19 and Pro $49 a month ($16 and $41 a month billed yearly). Teams of 3 or more: $17 and $42 per seat a month. (veridical.dev)Free covers public GitHub repositories, with no card, no application and no time limit, within fair-use limits shown in the app. For private repositories, choose Standard or Pro. (veridical.dev)Each finding names the affected code at the exact reviewed commit. When it can be checked, the repository's own tools run and the receipt is kept. (veridical.dev)The reference side of each comparison.

Every tool, A to Z

Order is not a ranking. Every fact links to the vendor's own page.

CodeRabbit

Best for
Teams on several Git hosts who want one reviewer on all of them.
Where it wins
Covers GitHub, GitLab, Azure DevOps and Bitbucket, and gives free reviews on public repositories with no application. (docs.coderabbit.ai)
Tradeoff
Priced per developer who opens pull requests, so cost grows with the number of authors. (coderabbit.ai)
Git hosts
GitHub, GitLab, Azure DevOps, Bitbucket Cloud and Bitbucket Data Center (docs.coderabbit.ai)
Price
Per developer: Essentials $24, Team $48, Advanced $72 a month billed yearly (Essentials $30 and Team $60 billed monthly). Enterprise on request. Only developers who open pull requests are charged. (coderabbit.ai, checked )
Evidence type
The model's review, plus output from 50+ linters and security analysers that “run automatically in secure sandboxed environments”. (docs.coderabbit.ai)

cubic

Best for
GitHub teams that want rules learned from past review comments and scheduled codebase scans.
Where it wins
Turns a team's past review comments into rules, and scans the whole codebase on a schedule. (cubic.dev)
Tradeoff
GitHub only, and usage is metered in reviewed lines per developer each month. (cubic.dev)
Git hosts
GitHub only (“We currently only support GitHub”) (docs.cubic.dev)
Price
Per developer: Team $40, Pro $99, Max $200 a month billed monthly, with a discount for yearly billing. Each plan includes a monthly allowance of reviewed lines per developer. (cubic.dev, checked )
Evidence type
The model's review against the whole codebase and the team's rules (“Every PR is checked against your whole codebase and your team’s rules”), plus a merge confidence score from 1 to 5. (cubic.dev)

Gemini Code Assist

Best for
Teams already on Google Cloud.
Where it wins
No charge while it is in Preview, and configuration across repositories from Google Cloud. (docs.cloud.google.com)
Tradeoff
Preview terms: pre-release features “might have limited support”. Setup needs a Google Cloud project. (docs.cloud.google.com)
Git hosts
GitHub, including GitHub Enterprise Server and Enterprise Cloud on the enterprise version (docs.cloud.google.com)
Price
Google's docs say “there are no charges for using Gemini Code Assist on GitHub during Preview”. A Google Cloud project with billing is required. (docs.cloud.google.com, checked )
Evidence type
The model's review, using information it “will automatically retrieve” from the repository and pull request, and an optional style guide. (docs.cloud.google.com)

GitHub Copilot code review

Best for
Organisations that already pay for Copilot.
Where it wins
Built into GitHub with no extra app, and it can submit an approving review when the organisation turns that on. (docs.github.com)
Tradeoff
Usage-based: each review draws on AI credits, and the agentic parts on Actions minutes. (docs.github.com)
Git hosts
GitHub.com, with Azure DevOps in public preview (docs.github.com)
Price
Included from Copilot Pro at $10 per user a month (Pro+ $39, Max $100). Reviews consume GitHub AI Credits, and agentic reviews also use GitHub Actions minutes. (github.com, checked )
Evidence type
The model's review with full project context gathering run on GitHub Actions. GitHub's own advice: “Always validate Copilot's feedback carefully.” (docs.github.com)

Greptile

Best for
Teams that want a reviewer to run the change, or that need to self-host.
Where it wins
TREX runs the changed code in a sandbox and attaches what it saw to the pull request, and Greptile can be self-hosted on several Git hosts. (greptile.com)
Tradeoff
Credits are priced by review depth, and TREX is still in beta. (greptile.com)
Git hosts
GitHub, GitLab and Bitbucket (cloud and self-managed), plus Gitea; self-hosting is offered (greptile.com)
Price
Starter is free for one developer. Pro is $1 per credit with no seat fees: a base review is 1 credit, Plus 3, Apex 10. Enterprise on request. (greptile.com, checked )
Evidence type
The model's review with codebase context. TREX, in beta, starts the services, exercises the changed code and “leaves the logs, screenshots, and traces on the PR”. (greptile.com)

OpenAI Codex code review

Best for
Teams already on a ChatGPT plan.
Where it wins
Comes with a ChatGPT plan and reads review rules from AGENTS.md. (learn.chatgpt.com)
Tradeoff
Flags only the highest-priority issues by design, so lower-priority problems are not raised. (learn.chatgpt.com)
Git hosts
GitHub (learn.chatgpt.com)
Price
Part of ChatGPT plans: Plus at $20 a month lists “Cloud-based integrations like automatic code review”; Pro from $100 a month. (learn.chatgpt.com, checked )
Evidence type
The model's review of the diff, following AGENTS.md guidance. On GitHub it “flags only P0 and P1 issues”. (learn.chatgpt.com)

Qodo

Best for
Larger teams with many repositories and shared rules.
Where it wins
Cross-repository review and a rules system, across GitHub, GitLab, Bitbucket and Azure DevOps. (qodo.ai)
Tradeoff
Usage is in credit packs, and the open-source plan needs 200+ stars. (docs.qodo.ai)
Git hosts
GitHub, GitLab, Bitbucket and Azure DevOps, plus Gerrit (docs.qodo.ai)
Price
Credit packs pooled across the team, from $30 a month on Pro Team; Enterprise on request. 14-day trial with no card. (qodo.ai, checked )
Evidence type
Multi-agent review against organisation rules, with cross-repository review (“Catch breaking changes before they reach other repos”). (qodo.ai)

Veridical

Best for
GitHub teams that want each finding backed by evidence they can check.
Where it wins
Findings carry their evidence. In our July 20, 2026 readout Veridical scored higher than CodeRabbit, cubic and Greptile on the credited-BT lens. (veridical.dev)
Tradeoff
GitHub only today. On the strict lens of the same readout, CodeRabbit, Gemini Code Assist and Greptile scored higher. (veridical.dev)
Git hosts
GitHub (github.com) only. GitLab and Bitbucket are not supported yet. (veridical.dev)
Price
Standard $19 and Pro $49 a month ($16 and $41 a month billed yearly). Teams of 3 or more: $17 and $42 per seat a month. (veridical.dev)
Evidence type
Each finding names the affected code at the exact reviewed commit. When it can be checked, the repository's own tools run and the receipt is kept. (veridical.dev)

Same-PR test: coming

We are setting up a public test repository with real, licence-checked pull requests. Each tool will run on its free or trial plan, and every comment will be marked Verified true, False or Could not verify, with a link to the evidence. This section stays empty until that repository is public: we do not show screenshots we did not take there. Until then, the benchmark below is the measured comparison.

When to keep the reviewer you have

  • It runs on the Git host you use and Veridical does not (GitLab, Bitbucket, Azure DevOps, Gitea).
  • Your team acts on most of its comments and rarely finds them wrong. That is what the pilot below measures.
  • It is already paid for inside a plan you keep for other reasons, such as Copilot or ChatGPT.

When not to pick Veridical

Veridical reviews pull requests on GitHub (github.com) only. If your code lives on GitLab, Bitbucket, Azure DevOps or a self-hosted Git server, Veridical is not a fit today; GitLab and Bitbucket are not supported yet.

A two-week pilot

  1. 01 Pick one active repository and install both your current reviewer and Veridical on it for the same two weeks.
  2. 02 Let both review the pull requests opened during the pilot. Do not tune either tool halfway through.
  3. 03 For each comment, record one of three outcomes: acted on (the code changed because of it), false (it was wrong once someone checked), or could not verify (nobody could settle it).
  4. 04 Time how many minutes the team spends reading and answering each tool's comments.
  5. 05 At the end, compare findings acted on, findings that were false, and minutes spent per finding acted on.

Questions

What is the best AI code review tool?
There is no single answer. Start from your Git host, then decide what you want behind each comment: a model's reading, linter output, a run of the changed code, or code cited at the exact commit. Then run two tools side by side on one repository for two weeks and count the findings that were acted on, the ones that were false, and the minutes spent reading.
How were the tools measured?
On real merged pull requests from public repositories where each tool had already commented. Veridical reviewed the same pull requests afterwards, and both sides were scored by the same rules on two lenses, credited-BT and strict, with a 95% confidence interval per tool. How many findings a run was attempted on differs by comparison and by side; the false-positive page gives the counts. The numbers come from our readout of July 20, 2026 (protocol 1.6.2).
Why are some tools marked not measured?
Our benchmark covers the tools whose review bots we sampled on public pull requests. GitHub Copilot code review and Qodo were not among them, so we compare them on their own published facts only.
Do these numbers describe today's Veridical?
No. The Veridical side of this run was our July 2026 research pipeline, which predates the production engine that reviews pull requests today. Every number is printed with its readout date for that reason.
Which tools support GitLab or Bitbucket?
From their own pages: CodeRabbit (also Azure DevOps), Greptile (also Gitea) and Qodo (also Azure DevOps). cubic and Veridical are GitHub only. GitHub Copilot code review lists Azure DevOps in public preview.
Do you show screenshots of other tools' comments?
Not yet. We are setting up a public test repository with real pull requests so that every screenshot comes from a run anyone can inspect, with each comment marked Verified true, False or Could not verify. Until then the pages show the benchmark only.

Put evidence on your next GitHub pull request.

You sign in with your GitHub account. The first time, GitHub asks you to authorize Veridical.

Get started