Preprint finds low-cost moderation model competitive, with confidence limits

Researchers report that a low-cost content moderation model achieved accuracy competitive with frontier reasoning models in a preprint audit covering 142,658 human-labeled comments. The model costs approximately USD 0.000025 per call and returns decisions in approximately 100 ms. On a shared probe set, its classification AUROC of 0.83 was not statistically distinguishable from 13 of 14 frontier systems while costing 1-3 orders of magnitude less. The comparison used n=56; failure to detect a difference does not establish equivalence. The audit also found that confidence scores conflate human disagreement with the risk of a costly wrong decision. Parallel sub-questions recovered disagreement signals for weaker models, but gains depended on the data distribution and how the questions were framed.