GPT-OSS 120B (Workers AI)

GPT-OSS 120B is a low-cost reviewer ($0.0075 and 55s per review on Workers AI) that is hard to recommend, because a closed model from the same maker beats it on recall, cost and noise.

That model is GPT-6 Luna. Luna measurably catches more bugs, its review costs less, and its false-alarm interval sits wholly below this model's. GPT-OSS 120B caught 40% [29%–52%] of the known bugs and raised 2.5 [2.2–2.9] false alarms per clean PR; Luna's figures are 67% [53%–80%] and 0.7 [0.3–1.1]. The side-by-side shows it fixture by fixture.

Look at where its comments land as well. 49% of its catches sat on the defect's own lines, fewer than Luna's 97%. When a finding points near the bug rather than at it, someone has to go looking, which eats into the time a cheap reviewer is meant to save.

Its rank is 8 of 17, with 7 ranked models measurably ahead, and it ran at medium reasoning effort. One scoring detail: an OpenAI judge sits on the three-judge panel, so its vote is dropped for this model.

The case for it is the open weights, and nothing else in this data. If you need an open model in a cheap seat, GLM-5.3 Flash cost less per review. GPT-OSS 120B on Groq lists at $0.15 / $0.60 per 1M tokens, against $0.35 / $0.75 here, but it was not in the run.