Abstract Structure-based methods for detecting protein-ligand binding sites play a crucial role in various domains, from fundamental research to biomedical applications. However, current prediction methodologies often rely on holo (ligand-bound) protein conformations for training and evaluation, overlooking the significance of the apo (ligand-free) states. This oversight is particularly problematic in the case of cryptic binding sites (CBSs) where holo-based assessment yields unrealistic performance expectations. To advance the development in this domain, we introduce CryptoBench, a benchmark dataset tailored for training and evaluating novel CBS prediction methodologies. CryptoBench is constructed upon a large collection of apo-holo protein pairs, grouped by UniProtID, clustered by sequence identity, and filtered to contain only structures with substantial structural change in the binding site. CryptoBench comprises 1,107 structures with predefined cross-validation splits, making it the most extensive CBS dataset to date. To establish a performance baseline, we measured the predictive power of sequence- and structure-based CBS residue prediction methods using the benchmark. We selected PocketMiner as the state-of-the-art representative of the structure-based methods for CBS detection, and P2Rank, a widely-used structure-based method for general binding site prediction that is not specifically tailored for cryptic sites. For sequence-based approaches, we trained a neural network to classify binding residues using protein language model embeddings. Our sequence-based approach outperformed PocketMiner and P2Rank across key metrics, including AUC, AUPRC, MCC, and F1 scores. These results provide baseline benchmark results for future CBS and potentially also non-CBS prediction endeavors, leveraging CryptoBench as the foundational platform for further advancements in the field.