aiqbench analysis

Coinbase reports that a post-trained open-weight 9B model outperforms Opus 4.5 on fraud detection, a narrow, task-specific result. For AI readers, the useful questions are how the comparison was measured (dataset, metrics, and whether the baseline was tuned) and whether a small specialized model can match a frontier model on a production-style task at lower cost and with more control. The article itself is thin, so treat the claim as unverified until the evaluation details are published.