Gemma 4 26B A4B (Workers AI)

Gemma 4 26B A4B rarely raises a false alarm, 0.4 [0.2–0.6] per clean PR, but it catches few bugs: 26% [15%–38%]. A reviewer this blind isn't worth a seat, however quiet it is.

The quiet is real, not a side effect of saying nothing. Its false-alarm interval sits wholly below Claude Haiku 5.5's and GLM-5.3 Flash's, and it still wrote long reviews, several times the output the spec sheet's review call assumes. The false positives write-up explains why a low alarm count only means something next to recall.

Rank 13 of 17 tells the rest: 12 ranked models measurably beat it, and it measurably beats 0. Claude Haiku 5.5 catches measurably more and its measured review costs less, even though a typical Haiku call lists at 1.2× a Gemma one. Gemma's long output more than erases that list-price edge. A review also took 308s.

For an open-weight alternative, GLM-5.3 Flash measurably catches more. Among Gemma's misses is "fix(saml): IdP-Initiated Callback Routing", which GLM-5.3 Flash's larger sibling caught every run.

A Google judge is one of the three, so its vote is dropped when scoring this model. The open-model results post covers the earlier open-weight runs.