Geospatial Foundation Models (GFMs) emerged as powerful tools for analyzing Earth observation data. Yet their evaluation often falls short, limiting the assessment of promising innovations. Many benchmarks rely on overly simplistic tasks and limited datasets, lacking diversity in resolution, sensor type, and global coverage. To this aim, we present PANGAEA, a comprehensive evaluation protocol that unifies diverse datasets, tasks, and sensors into a standardized benchmark. In addition, we show a couple of recent innovations in enhancing GFMs capabilities: DEFLECT, a parameter-efficient fine-tuning strategy based on embedding deflection, and SatDiFuser, a diffusion-based framework for tackling discriminative remote sensing tasks.