Abstract In fast vision, local spatial properties of the visual scene can automatically capture the observer’s attention. We used specific local features, predicted by a constrained maximum-entropy model to be optimal information-carriers, as candidate “salient features''. Previous studies showed that participants choose these optimal features as “more salient” if explicitly asked. Here, we investigated the implicit saliency of these optimal features in two attentional tasks. In a covert-attention experiment, we measured the luminance-contrast threshold for discriminating the orientation of a peripheral gabor. In a gaze-orienting experiment, we analyzed latency and direction of saccades towards a peripheral target. In both tasks, two brief peripheral cues, differing in saliency according to the model, preceded the target, presented on the same (valid trials) or the opposite side (invalid trials) of the optimal cue. Results showed reduced contrast thresholds, saccadic latencies, and direction errors in valid trials, and the opposite in invalid trials, compared to baseline values obtained with equally salient cues. Also, optimal features triggered more anticipatory saccades. Similar effects emerged in a luminance-control condition. Overall, in fast vision, optimal features automatically attract covert and overt attention, suggesting that saliency is determined by information maximization criteria coupled with computational limitations.