Specialization Beats Scale: 3B Model Outperforms Claude Opus 4.6 on Domain Task at 52x Lower Cost

Summary

Dharma-AI researchers Erick Lachmann and Pimenta de Freitas Cardoso published a quantitative case for specialization over scale in enterprise AI procurement. On a Brazilian Portuguese legal/administrative OCR benchmark (DharmaOCR), a 3-billion-parameter specialized model scored 0.911 composite vs. Claude Opus 4.6's 0.833 — an 8+ point gap — while costing approximately 52x less per million pages and showing 0.20% text degeneration vs. 1.41% for an unspecialized 3B baseline. The paper's framing is