The proliferation of fake news has been amplified by the advent of large language models (LLMs), which can generate highly realistic and scalable misinformation. While prior studies have focused primarily on detecting human-generated fake news, the efficacy of current models against LLM-generated content remains underexplored. We address this gap by compiling a novel dataset combining public and LLM-generated fake news, redefining detection as a ternary classification task (real, human-generated fake, LLM-generated fake), and evaluating eight diverse classification models, including traditional machine learning, fine-tuned transformers, and few-shot prompted LLMs. Our findings highlight the strengths and limitations of these models in detecting evolving LLM-generated fake news, offering insights for future detection strategies.