Cognitive deficits are present in the majority of psychosis patients. They range across various domains, such as working memory, and executive functioning, and are linked to neurobiological changes, including changes in gray matter volume. In this study, we explored data-driven clustering solutions using behavioural, demographic, psychological and structural brain data, but no clinical data, in a cross-diagnostic sample combining affective (N=51) and non-affective (N=111) psychosis patients as well as healthy controls (N=55). The goal of the study was (1) to test classification sensitivity of diagnostic groups using data without diagnostic information, and (2) to explore identification of potential cognitive phenotypes. We used K-means and spectral clustering, investigating two-, three- and four-cluster solutions, and group-membership matching. We then explored cognitive deficits and symptom expression within the four-cluster solution. Our results revealed best group-cluster matching using PCA-pre-selected only non-brain, mostly cognitive, features. The three groups were clustered with medium sensitivity, correctly identifying between 44% and 78% of individuals per group. More importantly, however, clustering using four clusters allowed the identification of cognitive phenotypes, that significantly varied in clinical and cognitive impairment. We identified one cluster expressing the lowest symptom scores and unimpaired cognition, one other cluster with the highest symptom scores, especially negative symptoms, and global cognitive impairments, across all domains, and two intermediate clusters. In conclusion, these results provide evidence for cognitive phenotypes with specific symptom expressions combining individuals with different overall diagnoses. The clear link between the cognitive deficits and symptoms indicates the need for the development of cognition-based interventions.