For the task of classifying news articles into their subject type (e.g., finance, science, politics), what is a current state-of-the-art approach? Describe this approach in detail and discuss its latency. Additionally, suggest simpler, classical models that can achieve faster throughput while maintaining high accuracy for this task.