Vision-language models (VLMs) have the potential to enhance machine translation (MT) by leveraging visual context alongside text, yet their real utility for production workflows remains unclear. We conduct a unified, multi-condition evaluation of six leading VLMs—both open and proprietary—on two challenging benchmarks (CoMMuTE and CaMMT), targeting lexical and cultural disambiguation respectively, with a domain-style case study simulating technical documentation localization. Results show that model performance varies widely, and the benefit of relevant images does not necessarily transfer across use cases. Proprietary models are notably sensitive to irrelevant images while open-source models are generally more stable; incorrect or contradicting visuals, by contrast, degrade translation across all models. Taken together, these findings make rigorous evaluation a necessary precondition for production deployment: metric gains can mask real accuracy losses in technical domains, model sensitivity to irrelevant images should inform model selection, and reliable image–text matching is a hard requirement for any pipeline.