Unpacking Political Bias in Large Language Models: A Cross-Model Comparison on U.S. Politics