Large language models (LLMs) and Neural Machine Translation (NMT) systems have demonstrated strong performance across a wide range of translation tasks. However, translating culturally grounded content such as idiomatic expressions remains a persistent challenge due to their non-compositional nature. This challenge is particularly pronounced in low-resource African languages such as Yoruba, where idioms carry deep cultural meaning yet remain underrepresented in machine translation research. In this study, we evaluate three models: GPT-4o mini, Qwen2.5-7B-Instruct, and NLLB-200 on Yoruba-to-English idiom translation. We examine direct machine translation as well as zero-shot and few-shot prompting strategies using a curated dataset of Yoruba idioms annotated with their literal meanings and corresponding English idiomatic equivalents. Evaluation is conducted using BLEU, chrF+, and BERTScore F1, alongside human evaluation based on adequacy, fluency, and preservation of literal meaning. The results show that GPT-4o mini performs best overall, particularly in the zero-shot and few-shot settings, according to the chrF+ metric. Notably, GPT-4o mini consistently achieves the highest scores for literal meaning across all models in human evaluation. This suggests a tendency for LLMs to default to literal translations when handling culturally grounded expressions. These findings highlight the limitations of current translation models in capturing figurative and pragmatic meaning, and underscore the need for culturally grounded corpora for low-resource African languages, particularly Yoruba, to support the preservation of cultural identity in language technologies.