Cryptocurrency markets are highly volatile and sensitive to public sentiment, making sentiment analysis of related news an essential tool for investors and analysts. In this paper, we evaluate the performance of large language models (LLMs), particularly GPT-4, on the task of cryptocurrency sentiment classification. We implement a domain-adaptive fine-tuning approach, where GPT-4 is instruction-tuned on a labeled dataset of crypto-related news headlines. This allows the model to better capture domain-specific linguistic cues and sentiment nuances. We benchmark GPT-4 against several state-of-the-art models, including Flan-T5, FinBERT, BERT, and Gemma-7B, using standard evaluation metrics and runtime analysis. Experimental results demonstrate that the fine-tuned GPT-4 achieves the most balanced performance across sentiment classes, while models like Flan-T5 offer faster inference with competitive accuracy. Our findings highlight the trade-offs between accuracy, efficiency, and deployability, offering practical insights for the application of LLMs in sentiment-driven financial decision-making.