A clinical benchmark of public self-supervised pathology foundation models is presented, evaluating the performance of various models on clinically relevant tasks. The study uses datasets from two medical centers, including clinical slides associated with cancer diagnoses and biomarkers. Public pathology foundation models, such as CTransPath, UNI, Virchow, and Prov-GigaPath, are assessed on tasks like disease detection, biomarker prediction, and treatment outcome prediction. The models are evaluated using a clinical benchmark dataset generated during standard hospital operations, covering a wide range of diseases and anatomical sites. The study compares the performance of different models, including those trained with DINO and DINOv2 algorithms, and finds that DINO and DINOv2 models generally outperform others. The results show that while larger models may perform better in some tasks, smaller models can also achieve comparable results. The study also highlights the importance of pretraining dataset composition, as models trained on datasets with more representation of certain tissues tend to perform better in tasks related to those tissues. The study also discusses the computational resources required to train these models, noting that DINO and DINOv2 are more computationally intensive than other algorithms. The results indicate that while larger models may offer some advantages, the benefits vary depending on the task and the pretraining dataset composition. The study concludes that while self-supervised learning and foundation models show promise in computational pathology, there is still a need for further research to improve their performance and applicability in clinical settings. The benchmark is made available for future research and development.A clinical benchmark of public self-supervised pathology foundation models is presented, evaluating the performance of various models on clinically relevant tasks. The study uses datasets from two medical centers, including clinical slides associated with cancer diagnoses and biomarkers. Public pathology foundation models, such as CTransPath, UNI, Virchow, and Prov-GigaPath, are assessed on tasks like disease detection, biomarker prediction, and treatment outcome prediction. The models are evaluated using a clinical benchmark dataset generated during standard hospital operations, covering a wide range of diseases and anatomical sites. The study compares the performance of different models, including those trained with DINO and DINOv2 algorithms, and finds that DINO and DINOv2 models generally outperform others. The results show that while larger models may perform better in some tasks, smaller models can also achieve comparable results. The study also highlights the importance of pretraining dataset composition, as models trained on datasets with more representation of certain tissues tend to perform better in tasks related to those tissues. The study also discusses the computational resources required to train these models, noting that DINO and DINOv2 are more computationally intensive than other algorithms. The results indicate that while larger models may offer some advantages, the benefits vary depending on the task and the pretraining dataset composition. The study concludes that while self-supervised learning and foundation models show promise in computational pathology, there is still a need for further research to improve their performance and applicability in clinical settings. The benchmark is made available for future research and development.