format stringclasses 8
values | characters int64 1.09k 4.68k | claude55_tokens int64 620 2.2k | claude55_vs_indented_json_pct int64 -66 20 | claude45_tokens int64 476 1.87k | claude45_vs_indented_json_pct int64 -71 15 | gemini_tokens int64 482 1.87k | gemini_vs_indented_json_pct int64 -70 15 | gpt_tokens int64 427 1.6k | gpt_vs_indented_json_pct int64 -68 20 |
|---|---|---|---|---|---|---|---|---|---|
tsv | 1,087 | 621 | -66 | 477 | -71 | 483 | -70 | 427 | -68 |
csv | 1,087 | 620 | -66 | 476 | -71 | 482 | -70 | 429 | -68 |
toon | 1,142 | 647 | -65 | 502 | -69 | 508 | -69 | 474 | -65 |
markdown_table | 1,453 | 800 | -57 | 649 | -60 | 590 | -64 | 539 | -60 |
json_minified | 2,447 | 1,210 | -34 | 847 | -48 | 791 | -51 | 763 | -43 |
yaml | 2,493 | 1,275 | -31 | 1,074 | -34 | 1,076 | -34 | 997 | -25 |
json_pretty | 4,230 | 1,841 | 0 | 1,618 | 0 | 1,623 | 0 | 1,336 | 0 |
xml | 4,675 | 2,200 | 20 | 1,868 | 15 | 1,873 | 15 | 1,603 | 20 |
Token cost by data format: JSON vs YAML vs CSV vs TOON on Claude, Gemini and GPT
The same 20-row product table (7 fields) written in 8 formats and counted on four tokenizers: Claude Opus/Sonnet 5.5, Claude Haiku 4.5, Gemini 3.x and OpenAI o200k_base (GPT-4o and later).
Headline: the table costs 1,336 GPT tokens as indented JSON and 429 as CSV (68% fewer). CSV was the cheapest format on every tokenizer tested; TOON came close but did not beat it; XML was the most expensive.
- Source page, chart and method: https://tokensave.app/data/token-cost-by-format
- Related: token cost by language on Claude, Gemini and GPT: https://tokensave.app/data/token-cost-by-language-claude-gemini
- DOI (Zenodo): 10.5281/zenodo.23435788
- Measured: 2026-10-17
- License: CC BY 4.0
Key findings
- Switching from indented JSON to CSV cuts tokens by 68% on GPT, 66% on Claude 5.5 and 70% on Gemini.
- TOON needs 10% more tokens than CSV on GPT and 4% more on Claude 5.5.
- Minifying JSON saves 51% on Gemini and 43% on GPT but only 34% on Claude 5.5.
- Claude 5.5 needs 28% to 59% more tokens than GPT for the same data (largest gap: minified JSON).
Files
token-cost-by-format.csv: one row per format with character count, tokens and difference vs indented JSON forclaude55,claude45,gemini,gpttoken-cost-by-format.json: the same results plus all 8 input files verbatim, so every count can be reproduced
Data
Tokens for the same table, difference versus indented JSON on the same tokenizer in brackets.
| Format | Claude Opus/Sonnet 5.5 | Claude Haiku 4.5 | Gemini 3.x | GPT o200k |
|---|---|---|---|---|
| TSV | 621 (-66%) | 477 (-71%) | 483 (-70%) | 427 (-68%) |
| CSV | 620 (-66%) | 476 (-71%) | 482 (-70%) | 429 (-68%) |
| TOON | 647 (-65%) | 502 (-69%) | 508 (-69%) | 474 (-65%) |
| Markdown table | 800 (-57%) | 649 (-60%) | 590 (-64%) | 539 (-60%) |
| JSON (minified) | 1,210 (-34%) | 847 (-48%) | 791 (-51%) | 763 (-43%) |
| YAML | 1,275 (-31%) | 1,074 (-34%) | 1,076 (-34%) | 997 (-25%) |
| JSON (indented) | 1,841 (±0%) | 1,618 (±0%) | 1,623 (±0%) | 1,336 (±0%) |
| XML | 2,200 (+20%) | 1,868 (+15%) | 1,873 (+15%) | 1,603 (+20%) |
Method
- One table of 20 products x 7 fields (id, name, category, price, in_stock, rating, tags), written as indented JSON (2 spaces), minified JSON, YAML, XML, CSV, TSV, a Markdown table and TOON (tabular form, tags joined with |).
- Claude: each file sent through Claude Code (
claude -p) with a fixed one-line system prompt; reported input tokens minus the same run with a one-character message. Opus 5.5 and Sonnet 5.5 gave identical counts. - Gemini: Google AI Studio prompt token counter on Gemini 3.8 Flash, minus its one start token. XML, CSV and indented JSON were measured twice; CSV and JSON matched, and the XML value is from the second run.
- GPT: tiktoken with
o200k_base.
Limitations
- One flat table. Deeply nested data does not fit CSV well; there JSON or YAML may be the practical choice.
- Token count only: models may read some formats more reliably than others.
- Claude values can be off by about ±2 tokens (under 1% here).
Citation
TokenSave (2026). Token cost by data format: JSON vs YAML vs CSV vs TOON on Claude, Gemini and GPT [Data set]. Zenodo. https://doi.org/10.5281/zenodo.23435788
@misc{tokensave2026format,
title = {Token cost by data format: JSON vs YAML vs CSV vs TOON on Claude, Gemini and GPT},
author = {TokenSave},
year = {2026},
publisher = {Zenodo},
doi = {10.5281/zenodo.23435788},
url = {https://tokensave.app/data/token-cost-by-format},
note = {Data set, CC BY 4.0}
}
Count tokens for your own data at tokensave.app.
- Downloads last month
- 33