TOON Libraries: Minimal Benchmark

This is a small, objective experiment that compares three TOON libraries (toonify, toons, toon-formatter) for payload size, token counts, and dumps/loads performance. It also validates typed round-trips with Pydantic. It runs locally with no network calls.

What it measures

• Bytes and token counts for a small RAG-style context payload.
• Bytes and token counts for a small answer payload.
• Dumps/loads timing across toonify, toons, and toon-formatter.
• Typed validation for TOON outputs using Pydantic models.

Setup

Install dependencies:

python -m pip install -r requirements.txt

Run

python run.py

Notes

• Token counts use tiktoken with gpt-5.2 as the reference tokenizer. Change MODEL_NAME in run.py if you want a different tokenizer.
• Iteration count is controlled by ITERATIONS in run.py.
• TOON output uses short keys to reflect the common production pattern where key shortening compounds savings.
• Cost estimation uses the PRICING table in run.py (USD per 1M tokens).