My Hobby Project's $847 API Bill: Why JSON is Burning Your Money (And When TOON Can Save You)
Lately, I've been obsessing over something that seemed trivial at first: data formatting for LLMs. I know—sounds about as exciting as watching paint dry. But here's what caught my attention: while building a hobby project—a personal knowledge base using RAG—my Claude API bill hit $847 in one month. For a hobby project. After one afternoon of testing different formats, I cut that by 42%.
The culprit? JSON. The hero? A format most people haven't heard of called TOON.
But here's where it gets interesting—and where the LinkedIn "hype train" gets it dangerously wrong. TOON isn't a silver bullet. I've also seen it increase costs by 20% and tank accuracy by 15% when used incorrectly.
This isn't about "JSON bad, TOON good." It's about understanding the hidden economics of the token-based economy we're all now operating in.
The Part That Made Sense
What struck me was this: We've moved from a world measured in bytes to a world measured in tokens. And in that shift, JSON transformed from an efficient format into a financial liability.
Every brace, bracket, comma, and quote in your JSON is a billable item. But the real killer? Repetition.
Imagine you're sending a list of 1,000 user records to an LLM. In JSON, you repeat the keys "id," "name," and "role" exactly 1,000 times:
{
"users": [
{"id": 1, "name": "Alice", "role": "admin"},
{"id": 2, "name": "Bob", "role": "user"},
// ... 998 more times
]That's not just inefficient—it's financially reckless. At current pricing (Claude 3.5 Sonnet: $3 per million input tokens), this redundancy is literally burning your budget.
The cost isn't just financial. JSON creates three compounding problems in the LLM world: