Updated 17 August 2026

How to choose an AI code reviewer.

Three vendors currently claim to be number one on the same benchmark. All three are telling the truth. Here is what the numbers mean, why they disagree, and how to get an answer about your own codebase instead.

We build one of these tools, so read this with that in mind. What follows contains no claim that ours is better — we have not run that comparison, and neither has anyone else in a way that would settle it. What it does contain is every published number we could source, with the date attached, because the date is the part everyone leaves out.

There is finally an independent benchmark

Until 2026 the only comparisons were vendor blog posts. Then Martian — a lab that does not sell a review tool — published Code Review Bench: seventeen reviewers measured against real pull requests from open-source repositories, with the dataset, judge prompts and pipeline open sourced.

It reports three things, and the definitions matter more than the scores:

MetricWhat it actually asks
PrecisionOf the comments the tool posted, what share matched a change the developer went on to make?
RecallOf the fixes the developer made after review, what share had the tool already flagged?
F1The balance between the two.

Note what precision is not: it is not "was the comment correct". A comment can be technically right and still fail, because the developer read it and decided it did not matter. That is the correct bar. A reviewer that is right about things you will not act on is still wasting your attention.

Why every vendor claims first place

The leaderboard updates continuously as new pull requests are analysed. So the ranking on any given week is a snapshot, and each vendor published the week it won:

ToolPublished claimSnapshot
cubic 61.8% F1, stated as #1 2026
Greptile 60.8% F1, 76.2% precision, 50.6% recall — highest precision measured 30 July 2026
CodeRabbit 51.2% F1, highest recall of any tool by roughly 15 points Jan–Feb 2026 window
CodeAnt 51.7% F1, reported third overall 2026
Do not read this as a ranking

These figures come from different dates against a moving dataset. Putting them in one column and sorting it would be the same trick each vendor is playing, with extra steps. Use the table to see the range and the shape, not to pick a winner.

What the range tells you

The useful reading is not who leads. It is where the ceiling is.

The best precision anyone has published is 76.2%. At that level, roughly one comment in four still does not correspond to anything the developer changed — and that is the best measured, not the average. Further down the same benchmark tools land near 47% precision, where more than half the output is noise.

Recall at the top sits near 50%. Whichever tool you choose, plan on it catching about half of what your developers will end up fixing. None of these replaces review; they change what a human is spending attention on.

Precision and recall pull in opposite directions

Notice that the two most-cited claims are on opposite axes. Greptile leads on precision — the quietest, least wasteful output. CodeRabbit led on recall — catching the most. Neither is the better tool in the abstract, and a tool that led on both would be suspicious.

If you optimise forYou getSuits
PrecisionFewer comments, more of them worth acting on, more missesTeams who have already been burned and stopped reading the bot
RecallMore real issues caught, more noise to wade throughSecurity-sensitive paths where a miss costs more than an interruption

Most teams say precision and then configure for recall.

The axis the benchmark does not measure

Every metric above is per comment. Whether a tool survives in your repository is per inbox, and it depends on volume as much as accuracy.

CodeAnt's analysis of teams running these at scale describes 200–400 AI comments a week with 70–90% dismissed. At that volume people do not evaluate each comment and reject the bad ones — they stop reading the column. A quieter tool at the same precision can be worth more than a louder one, because it is still being read in month six.

No benchmark captures this, because it happens after the measurement window. It is the single best reason to run your own trial rather than buy from a leaderboard. We wrote about the mechanism behind it in why teams mute their AI code reviewer.

What they cost

List prices as published, per developer per month. Check before buying — these move.

ToolPriceNote
CodeRabbit~$24 annual, $30 monthlyPro Plus around $48 adds test generation
Greptile$30 per seatIncludes 50 reviews; overage charged per review
Qodo$30 per user
Cursor Bugbot~$40
GitHub Copilot code reviewBundled with CopilotEffectively free where you already pay for Copilot

The bundled option matters more than its quality suggests. If your team already has Copilot, the baseline is not "no reviewer" — it is a reviewer you are already paying for. Anything else has to beat that, not zero.

Run the comparison yourself

Two weeks, and it answers the question the benchmark cannot:

  1. Install two tools on the same repository. They post alongside each other. Do not tune either one yet.
  2. Count comments posted, per tool, per week. Raw volume.
  3. Count comments acted on. A comment counts if it caused a code change or a real discussion. A thumbs-up does not. "Good catch, but no" does not.
  4. Watch the ratio's direction, not its value. Falling week over week means your team is disengaging, and that happens quietly.
  5. Ask in standup: did you read the bot on your last PR? The honest answer arrives faster than any dashboard.

Above roughly one in three acted on, a reviewer is earning its place. Heading toward one in ten, you are paying for something already ignored — and switching vendors buys a few weeks before the curve repeats.

Where Dicast fits

We are building around the observation in the section above: that the binding constraint is attention, not capability. A cheap wide scan proposes findings, then three jurors from three different model vendors — none of which wrote the finding, each asked to refute rather than confirm it — vote on whether it reaches your pull request. Two votes to post, three to block a merge. Everything struck stays on the record with the reason.

We are not on the benchmark and we are not claiming a number. When we have one measured the same way as everyone above, it will be on this page with its date attached, including if it is worse. Until then, the honest summary is that we think the industry is optimising the wrong axis, and we are finding out whether anyone agrees.

If you are choosing right now

Pick from the table above on precision or recall depending on which mistake costs you more, run the two-week test in your own repository, and ignore anyone claiming a leaderboard position without a date next to it — ourselves included.