Current language models struggle with context learning, which is essential for solving complex real-world tasks. The article highlights three scenarios to illustrate this challenge: converting flight requests into SDK pseudocode, learning a new game from its rulebook, and analyzing experiment logs to derive relationships. Language models primarily rely on pre-trained knowledge and fail to adapt or learn from new context during inference. This creates a gap between their capabilities and user needs, as they often overlook critical contextual details or revert to prior assumptions. Evaluation of ten state-of-the-art language models on the CL-bench shows that they solve only 17.2% of tasks, with the best-performing model achieving 23.7%. A significant reason for these failures lies in how models handle context. Many errors arise from misusing or ignoring context rather than lacking information. Even when models can process long inputs, they still often struggle with tasks that require tracking complex dependencies or following new instructions. The complexity of context, not just its length, contributes to these challenges. CL-bench was designed to focus specifically on tasks requiring models to learn new information from context. Each task is self-contained, removing the possibility of relying on external knowledge or assumptions. This approach aims to ensure that performance reflects genuine context learning rather than mere memorization. The findings point to a critical gap in current language models' abilities to learn from context, which affects their effectiveness in practical applications.