TOON alternatives: other ways to cut JSON tokens in LLM prompts
Updated by json2toon.co
The main alternatives to TOON for cutting LLM prompt tokens are minified JSON, CSV, YAML, TONL, TRON and JTON, plus the option of simply sending less data. Minified JSON wins on small or deeply nested payloads and CSV on a single flat table. TOON wins on arrays of similar records with structure around them. TRON and JTON are newer and have much less tooling.
Token counts on the same data
| Sample data | JSON | JSON (minified) | TOON | TONL | YAML | CSV |
|---|---|---|---|---|---|---|
| GitHub issues: Flat uniform array: 8 records, 7 fields each | 587 | 388 (-34%) | 294 (-50%) | 302 (-49%) | 463 (-21%) | 273 (-53%) |
| Stripe charges list: Object wrapping a uniform array: 7 records, 7 fields each | 512 | 335 (-35%) | 246 (-52%) | 271 (-47%) | 391 (-24%) | not possible |
| Nested orders: 3 records, each with a nested customer object and an items array | 219 | 122 (-44%) | 146 (-33%) | 201 (-8%) | 150 (-32%) | not possible |
| Kubernetes Deployment: Deeply nested config, no uniform arrays | 230 | 119 (-48%) | 132 (-43%) | 204 (-11%) | 132 (-43%) | not possible |
Which alternatives can be measured on the same data?
Five of them, because this site's converter produces all five. The table above shows each one on the same four samples. The short version:
- Minified JSON. Free, lossless and understood by everything. It beat TOON on both nested samples (122 vs 146 tokens and 119 vs 132) and lost on both tabular ones. If you send pretty-printed JSON today, minify it first. See TOON vs minified JSON.
- CSV. The smallest option for one flat table (273 tokens against 294 for TOON), and not an option for anything nested. See TOON vs CSV.
- YAML. No braces or quotes, but it still repeats every key for every list item, so it sits between JSON and TOON on tabular data and level with TOON on nested config. See TOON vs YAML.
- TONL. Close to TOON on tables and larger on nested data, in exchange for a query API, schema validation and streaming. See TOON vs TONL.
- TOON. For reference, TOON uses about 43% fewer tokens than JSON on average (42.6% in the official TOON v4.1 benchmark, 5,856 LLM calls across four models).
What about TRON and JTON?
This site does not convert to either, so the figures below come from their own sources and are not comparable with the table above or with each other.
| TRON | JTON | |
|---|---|---|
| Idea | Declare an object shape once as a class, then write instances inline | A strict JSON superset whose "Zen Grid" puts column headers in one row |
| Reported saving | Up to 27% fewer tokens in agent benchmarks | 15 to 60% fewer than compact JSON, 28.5% on average |
| Source | Notation Matters (arXiv 2605.29676) | JTON paper (arXiv 2604.05865) |
| Implementation | @tron-format/tron on npm | Rust/PyO3 reference described in the paper |
The same agent study measured TOON at up to 18% fewer tokens with about a 9 point accuracy cost on open-weight models, so in tool-calling loops TRON came out ahead of TOON on token reduction. Our TOON vs TRON vs JTON article goes through the syntax, the studies and the maturity of each project.
Is a different format even the right fix?
Often the largest saving is sending fewer fields. Dropping columns the model does not need shrinks every format at once. You can cut a payload down with a JSONPath query in the JSON tools and then encode what is left. Provider features stack with any format too: see prompt caching and the OpenAI Batch API.
How should you choose?
| Your data | Start with |
|---|---|
| One flat table | CSV, or TOON if types and a row count matter |
| Lists of similar records inside other structure | TOON |
| Small or deeply nested objects | Minified JSON |
| Config files | Leave as YAML or minify the JSON |
| A data file you query, validate or stream | TONL |
| Anything the model must write | JSON with structured outputs |
Try it on your data
Everything runs in your browser, so nothing you paste is uploaded. Convert a real payload, then check the token counts side by side before you decide.
Frequently asked questions
What is the best alternative to TOON?
It depends on the data shape. Minified JSON is the best alternative for small or deeply nested payloads, CSV for a single flat table, and TONL when you need query, schema or streaming tooling. For arrays of similar records with structure around them, TOON is usually the smallest of the formats we can measure.
Is minified JSON good enough instead of TOON?
For nested data, often yes: minified JSON used fewer tokens than TOON on our nested orders sample (122 vs 146). For tabular data TOON was 24% smaller than minified JSON on a flat array of 8 records.
Is TRON better than TOON?
In one 2026 agent study (arXiv 2605.29676) TRON reduced tokens by up to 27% against TOON's 18% in tool-calling workloads on open-weight models. TRON has far less tooling and adoption, and the official TOON benchmark covers a different task (retrieval on hosted models), so measure both on your own workload.
Can I reduce tokens without changing format?
Yes. Remove fields the model does not need, minify the JSON, and use provider features such as prompt caching and batch pricing. These stack with a format change.
Keep reading
- TOON vs TRON vs JTON
- TOON vs JSON vs YAML vs TOML: decision matrix
- Markdown tables vs TOON
- NDJSON vs JSON vs TOON
- Reducing LLM API costs