Gemini 3.7 Flash finds 11.6% of the bugs human reviewers caught. When it does report something, it is right 73.9% of the time. Both numbers are the extreme of the board: the lowest recall we measured and the highest precision, by a wide margin on each.
Read as a ranking, that model is last. Read as a posture, it is doing something the others are not.
Two ways to be wrong
A code reviewer can fail in two directions. It can stay quiet about a real problem, or it can raise something that is not one. Recall measures the first, precision the second, and no model on this board is good at both.
| Model | Recall | Precision |
|---|---|---|
| DeepSeek V4 Pro | 44.2% | 43.6% |
| Kimi K2.7 Code | 37.9% | 50.0% |
| GLM 5.2 | 35.2% | 47.0% |
| Muse Spark 1.2 | 41.0% | 35.0% |
| Gemini 3.7 Flash | 11.6% | 73.9% |
The model with the best recall is close to the middle on precision. The model with the best precision is last on recall. That trade is the shape of the whole board.
Which one you want is not a preference
It depends on what happens when the review fires.
If the review blocks a merge, precision is the cost that lands on people. A wrong block costs a developer twenty minutes and some goodwill, and it happens on every false positive. A team that stops trusting the gate routes around it, and then recall stops mattering because nobody is reading the output.
If the review is advisory, running beside a human who reads the diff anyway, the calculus flips. A missed security regression is the expensive outcome, and a noisy suggestion costs a scroll.
Why one number cannot carry this
F1 combines the two into a single figure, and we publish it because ranking needs a scalar. It also hides exactly the distinction above. Gemini 3.7 Flash has an F1 of 20.0, which reads as "bad", when what it actually is is "extremely conservative".
Nothing in the aggregate tells you whether a model is quiet and accurate or loud and thorough. The precision column does, and it is the one to read second.
The measurement caveat
Precision here is scored against the same golden set as recall: a finding counts as correct if it matches a bug a human reviewer reported on that pull request. A model that finds a genuine problem the humans did not mention is scored as a false positive. That penalises thorough models and flatters conservative ones, and it is a known limitation of using human review as ground truth.