TOON: Cutting LLM Token Costs by 40–60%

Token costs are the quiet tax of the AI era. Every time we send JSON to an LLM, we're paying for the same field names again and again. A 100-row dataset? You've basically paid for customer_id 100 times.

I stumbled upon TOON (Token-Oriented Object Notation) this week, and it's such a clever idea. Instead of repeating keys, you define the schema once, then stream your data like CSV:

orders[100]{id,customer,total}:
 1,John,499.00
 2,Ana,129.00

The benchmarks show 40–60% fewer tokens on real datasets — one analytics dump went from 10,977 → 4,507 tokens. Even better, models actually parse TOON more accurately than JSON, because the structure helps them "see" the data clearly.

If you're running anything that pushes structured data into LLMs, this is worth a look.