Accurate prediction of chemical properties and toxicological profiles is a critical challenge in cheminformatics and drug discovery. The Simplified Molecular Input Line Entry System (SMILES) provides a textual representation of molecular structures, enabling machine learning models to analyze large-scale datasets efficiently. Deep learning models, such as graph neural networks and sequence models, have demonstrated powerful capabilities in extracting features from SMILES and improving accuracy across various molecular prediction tasks. This study leverages advanced large language models (LLMs), including GPT-3.5 Turbo, GPT-4o, Gemini 1.5 Flash, and Gemini 1.5 pro, to perform zero-shot and few-shot learning on SMILES data. By comparing these approaches to traditional deep learning methods and prior work, we explore their potential in capturing contextual and sequential information for chemical property prediction. Our experiments demonstrate that zero-shot and few-shot learning enable high predictive accuracy in scenarios with limited labeled data, providing significant improvements over earlier methods. By systematically evaluating the performance of different LLMs on toxicity and physicochemical datasets, we highlight their strengths in enhancing prediction efficiency while addressing key challenges, such as dataset variability and interpretability. This work underscores the potential of LLMs to transform cheminformatics by offering scalable and versatile tools for molecular property prediction and accelerating advancements in drug discovery.