For vulnerability research, smaller models run repeatedly can outperform larger frontier models on cost-to-recall.