Artificial intelligence (AI) is increasingly embedded in digital assessment systems to support the synthesis of written feedback in higher education. While prior research has examined the reliability and structure of AI-generated feedback, considerably less attention has been paid to the pedagogical trade-offs between feedback fidelity, consistency, and environmental cost. This study investigates how different large language models (LLMs) vary in their capacity to generate structured, behaviourally anchored feedback in Objective Structured Clinical Examinations (OSCEs), and how these differences relate to relative environmental burden inferred from output volume. Using anonymised OSCE performance data from Qpercom, structured written feedback was generated for 51 stratified student profiles, and outputs were analysed as deployable feedback artefacts by quantifying verbosity and stage stability (internal examiner feedback to student-facing feedback drift and within-student variability), generation-time feasibility, and scenario-based emissions. The findings demonstrate substantial variation between models in verbosity, structural stability, and consistency across performance levels. One model produced substantially longer outputs with strong structural adherence but a markedly higher estimated environmental footprint, while another delivered more concise feedback with comparable pedagogical alignment and lower inferred emissions. These differences were driven primarily by output volume rather than assumed computational efficiency. AI-assisted feedback can enhance the structural quality and consistency of assessment narratives, but model selection and output governance materially affect both pedagogical coherence and sustainability. Rather than maximising feedback length, responsible educational deployment of AI requires explicit design constraints that balance fidelity, equity, and environmental considerations.