Intermediate ConvRNN circuits with highest OST consistency conserve on network size while maintaining task performance. Across all models considered, the intermediate ConvRNNs (denoted by $×$⁠) that attain high categorization performance (⁠$x$-axis) while maintaining a low unit count (panel B) rather than parameter count (panel A) for their given performance level, achieve the highest mean OST consistency (Spearman correlation with IT population OST, averaged across $N=10$ train/test splits). The color bar indicates this mean OST consistency (monotonically increasing from purple to red), binned into six equal ranges. Models with a larger network size at a fixed performance level are less consistent with primate object recognition behavior (e.g., deep feedforward models, denoted by boxes), with recurrence maintaining a fundamental trade-off between network size and task performance.