JSON input
tokens (est.)
—
TOON output
tokens (est.)
—
—
fewer tokens than JSON
Saved — tokens · JSON — → TOON —
Try a sample payload
Why TOON saves tokens
- Headers, not key repetition. A list of 1,000 user records repeats
"id","name","email"a thousand times in JSON. TOON declares the columns once and writes each row as a CSV-like line. - No quoted scalars. Identifiers (e.g. role names, status codes) don't need quotes in TOON, dropping two tokens per value.
- Structural punctuation collapses.
{}and[]become indentation.:and,stay but density drops.
The token estimator here is a heuristic (~5–10% off real GPT/Claude tokenizers). The relative comparison (TOON vs JSON for the same payload) is meaningful; the absolute counts shouldn't be your billing forecast.