We benchmarked coding agents on real engineering tasks across Databricks' multi-million line codebase to better understand the relationship between cost and performance. The benchmark surfaced several key findings. Top-tier performance now comes from a mix of proprietary and
