The GeoNatureAgent Benchmark introduces a new standard for evaluating AI agents in environmental geospatial analysis. This benchmark assesses seven large language models (LLMs) on their ability to perform 93 tasks using a real geospatial API. The findings highlight the cost-effectiveness of open-weight models and reveal significant limitations in reasoning capabilities. This advancement could streamline data analysis for environmental scientists, reducing time spent on data wrangling.