Kimi K2.7 Code (Workers AI)

Kimi K2.7 Code makes you wait for little: a review took 610s, cost $0.0567, and caught 38% [28%–47%] of the known bugs. There is no review seat where it is the right pick on this data.

It ran at high reasoning effort and wrote long answers, several times the output the spec sheet's review call assumes. None of that turned into recall. Both GLM models on Workers AI measurably catch more, and a Kimi review cost 16× a GLM-5.3 Flash review. The GLM-5.3 Flash comparison puts the two side by side.

Its results also wobble between runs. Some of the bugs it found, it found only sometimes; "caddyhttp: Optimize large host matchers" is one. The bug table's caught-every-run and caught-sometimes counts show how much of its recall is repeatable.

On noise it averaged 2.6 [2.0–3.2] false alarms per clean PR, and GPT-6 Luna's interval sits wholly below that. Luna also measurably beats it on recall, as the Luna comparison shows. Overall it ranks 11 of 17, and 10 ranked models measurably beat it.

Two sibling listings exist, Kimi K2.7 Code from Moonshot at $0.95 / $4.00 per 1M tokens and the Highspeed variant at $1.90 / $8.00; neither was in the run, so nothing here says how they review. The weights are open, but if that is the requirement, GLM-5.3 Flash meets it with more recall for less.