Though numerous studies have probed LLMs’ bias in political discourse, little is known about how their predictive patterns vary across partisan contexts, particularly under media influence. Leveraging GPT-4o as a predictive simulator for the 2024 U.S. Presidential election, the present work examines: a) the model’s fidelity in emulating electoral outcomes; b) the emergence of predictive bias across left- and right-leaning states; and c) the extent to which media exposure shapes forecasts across ideological lines. While the predicted outcomes strongly aligned with actual results, an asymmetric bias in forecasts was observed. GPT predictions consistently overestimated Democratic vote shares and underestimated Republican performance. This distortion remained most persistent in right-leaning states and tended to be more flexibly recalibrated within left-leaning territories. Swing states, despite comparatively lower bias, frequently displayed predictive misclassification. Given the growing integration of LLMs into political spheres, these findings offer nuanced insights into GPT-4o’s predictive capabilities across partisan alignments, shedding light on the ways in which the model aligns with observed realities or produces bias throughout electoral battlegrounds.