GLM-5.3 Flash (Workers AI)

If you want an open-weight model reviewing your pull requests, start with GLM-5.3 Flash: it caught 61% [50%–72%] of the known bugs for $0.0035 a review. Neither its larger sibling GLM-5.3 nor DeepSeek V4 Pro measurably beats it.

It ranks 2 of 17, and Claude Opus 5.5 is measurably ahead of it. Against the closed models the picture is closer than the rank suggests. GPT-6.1 Sol, GPT-6 Luna and Claude Haiku 5.5 can't be told apart from it on recall, and its review costs less than Haiku's.

Luna still wins on the things you feel day to day. Its review is cheaper, it raised fewer false alarms (0.7 [0.3–1.1] per clean PR against 1.8 [1.6–2.0], intervals apart), and it answered faster: 29s where GLM-5.3 Flash took 254s. So the case for GLM-5.3 Flash rests on its weights being open. If that matters to you, it gives up little on recall for it; if it doesn't, the head-to-head with Luna shows what you would trade.

Inside the family the choice is easy. A GLM-5.3 review cost 8.8× a Flash review, with no measurable recall gap to show for it; the family comparison lays out both. The open-model results post has background on how the Workers AI models were first run.