This paper studies whether large language models (LLMs) express systematically different opinions on contested questions in development economics and political economy, and whether those opinions are stable across languages. We ran a bilingual questionnaire of 83 multiple-choice questions on inequality, institutions, redistribution, labor markets, trade, and development focused on Latin America utilizing a set of widely used open-weight models under a common, reproducible inference pipeline. The LLM responses exhibit structured opinion profiles: a small number of latent dimensions explain most of the variation in outputs, and these dimensions correspond to recognizable political-economic interpretations rather than isolated idiosyncrasies. Additionally, these profiles are only partly language-invariant: the main structure of disagreement is broadly stable across English and Spanish, but language still affects secondary dimensions and question-level lead choices in a non-trivial share of cases. Because the design relies on repeated querying across multiple temperatures, the results also show that this structure is not reducible to one-off sampling noise. The findings suggest that LLM outputs on normative questions in (Latin American) political economy should be treated as structured, model-dependent judgments rather than as neutral defaults. JEL Classification: N4 , O1 , Z0