Скачать или смотреть JSON vs TOON

LLMs are powerful, but the moment you feed them large JSON payloads, the token bill starts to hurt. TOON solves that problem in a surprisingly elegant way.

Token-Oriented Object Notation is a compact, human-readable format that represents the same structures as JSON, but in a syntax that removes unnecessary punctuation and uses indentation the way YAML does. The real power shows up when you have uniform arrays of objects: instead of repeating keys on every row, TOON lets you declare the fields once and stream the data like a table.

For LLMs, this structure is ideal. Models can parse, validate, and retrieve information more reliably, while you save 30–60 percent tokens compared to formatted JSON in large array-heavy datasets. Benchmarks show that TOON delivers higher retrieval accuracy per token across multiple models, making it a strong choice when you’re pushing large inputs into prompts.

It’s not a replacement for JSON in your codebase. Think of it as the translation layer: keep JSON in your pipeline, but encode it as TOON when sending to an LLM. You get token efficiency without losing structure.

If your workflows involve agents, batch queries, analytics data, logs, or any uniform data streams, TOON gives you a predictable and compact format that models handle exceptionally well.