Structured Context Engineering Evaluates LLM SQL Performance

Damon McMillan publishes a new paper presenting 9,649 experiments on context engineering for structured data, evaluating 11 models and four formats with SQL schemas from 10 to 10,000 tables. The study finds frontier models (Opus 4.5, GPT-5.2, Gemini 2.5 Pro) outperform leading open-source models, and identifies a 'grep tax' for TOON formats increasing token usage. Results inform file-native agent design.