Verified Performance

Primary benchmark: 1,000 user records with tiktoken cl100k_base (GPT-4/Claude tokenizer).

JSON (minified)
24,005 tokens baseline
TOON
12,002 tokens -50.0%
Tauq SOTA
11,012 tokens -54.1%
12,993
Tokens Saved vs JSON
10.8%
Better than TOON (avg)
$0.04
Saved per 1K Records

* All measurements verified with tiktoken cl100k_base tokenizer
** Cost based on GPT-4o/Claude 3.5 Sonnet pricing ($3 per 1M input tokens)
*** TOON encoded via toon-python library for fair comparison. Tauq v0.1.0

Benchmark Methodology

Tokenizer

  • • tiktoken cl100k_base
  • • Used by GPT-4, GPT-4o, and Claude
  • • Industry-standard BPE tokenization
  • • Reproducible with Python tiktoken library

Fair Comparison

  • • TOON via toon-python library
  • • Proper [N]{fields}: headers
  • • Comma-delimited values per spec
  • • No artificial handicaps

Complete Results

We tested across 9 different dataset types to ensure a comprehensive comparison. All token counts verified with tiktoken cl100k_base. Results vary by dataset - flat structured data shows the highest efficiency gains (up to 54% vs JSON), while nested data shows more modest improvements (aggregate: 44% vs JSON, 11% vs TOON).

Dataset Tauq vs JSON JSON Tauq TOON Tauq vs TOON Winner
flat_100
100 user records (5 fields)
 -53.8% 2,402 1,109 1,199 -7.5% Tauq
flat_1000
1,000 user records (5 fields)
 -54.1% 24,005 11,012 12,002 -8.2% Tauq
mixed_structure
Nested objects with arrays
 -41.2% 689 405 457 -11.4% Tauq
deeply_nested
10 orgs with deep nesting
 -23.3% 6,990 5,359 7,630 -29.8% Tauq
wide_records
100 records with 15 fields each
 -53.1% 10,494 4,925 4,891 +0.7% TOON
heterogeneous
100 records with varying schemas
 -18.5% 1,594 1,299 1,765 -26.4% Tauq
timeseries
200 timestamp/value pairs
 -19.9% 5,003 4,007 3,799 +5.5% TOON
ecommerce
Product catalog with nested data
 -42.1% 2,970 1,719 1,879 -8.5% Tauq
api_response
Paginated API response
 -30.0% 1,089 762 706 +7.9% TOON
config_style
Realistic application config
 15.6% 411 475 502 -5.4% Tauq
TOTAL -44.2% 55,647 31,072 34,830 -10.8% Tauq

Tauq wins 7 of 10 dataset types. TOON wins on wide_records, timeseries, and api_response.

Performance by Category

-8.2%
Flat Tabular Data
TOON's primary use case
-26.4%
Heterogeneous Data
100 records with varying schemas
-29.8%
Deeply Nested
10 orgs with deep nesting

Primary Benchmark: 1,000 Records

Results Summary

Format Tokens vs JSON vs TOON Cost Savings*
JSON 24,005 baseline +100% $0.00
TOON 12,002 -50.0% baseline $0.036
Tauq 11,012 -54.1% -8.2% $0.039

* Cost based on GPT-4o/Claude Sonnet 4 pricing ($3 per 1M input tokens, Nov 2025)

** TOON encoded via toon-python library for fair comparison. Tauq v0.1.0

Why Tauq Beats TOON

1. Space Delimiters Beat Commas

In cl100k_base, spaces are often absorbed into preceding tokens, while commas create separate tokens.

TOON (comma-delimited)
1,User1,user1@example.com,21,false
Tauq (space-delimited)
1 User1 user1@example.com 21 false

2. Simpler Schema Syntax

Tauq's !def is more compact than TOON's count-prefixed headers.

TOON: Inline header with count
[1000]{id,name,email,age,active}:
Tauq: Clean schema definition
!def Record id name email age active

3. Bareword-Friendly Values

Common patterns like emails and paths don't need quotes in Tauq.

JSON: 3 tokens per quoted string
"user@example.com"
Tauq: 1 token bareword
user@example.com

4. No Count Prefix Required

TOON requires knowing array length upfront. Tauq supports true streaming.

Real-World Impact

$39-$65
Saved per 1M records*
12,993
Fewer tokens per 1K records
54%
Reduction vs JSON
10.8%
More efficient than TOON

* Based on GPT-4o/Claude Sonnet ($3/1M) to Claude Opus 4.5 ($5/1M) pricing as of Nov 2025

Sample Output Comparison

Here's how the same data looks in each format (first 3 records of flat_100):

JSON (minified)

[{"id":1,"name":"User1","email":"user1@example.com","age":21,"active":false},{"id":2,"name":"User2","email":"user2@example.com","age":22,"active":true},{"id":3,"name":"User3","email":"user3@example.com","age":23,"active":false}]
~72 tokens for 3 records

TOON (v3.0 spec)

[3]{id,name,email,age,active}:
  1,User1,user1@example.com,21,false
  2,User2,user2@example.com,22,true
  3,User3,user3@example.com,23,false
~42 tokens for 3 records

Tauq

!def Record id name email age active
1 User1 user1@example.com 21 false
2 User2 user2@example.com 22 true
3 User3 user3@example.com 23 false
~33 tokens for 3 records

Transparency Notes

  • • TOON wins in 3 scenarios: wide_records (+0.7%), timeseries (+5.5%), and api_response (+7.9%) where TOON's compact inline syntax helps.
  • • Tauq's advantage comes from: space delimiters (absorbed into tokens), bareword emails/paths, and simpler schema syntax.
  • • TOON encoded fairly: We use the official toon-python library, not a strawman implementation.
  • • Reproducible: Run docker build -t tauq-benchmark . && docker run --rm tauq-benchmark in the benchmarks directory to verify all numbers.

Run Your Own Benchmarks

Verify these results yourself using our benchmark suite: 

git clone https://github.com/epistates/tauq
cd tauq/benchmarks
docker build -t tauq-benchmark .
docker run --rm tauq-benchmark

The benchmark generates all test datasets, encodes them in JSON/TOON/Tauq, and counts tokens with tiktoken cl100k_base.

Feature Comparison

Feature Tauq TOON JSON
Token efficiency (flat data) -54% vs JSON -50% vs JSON baseline
Streaming support Native (no count prefix) Requires [N] prefix SAX/streaming parsers
Schema reuse across document Yes (!def/!use) No (inline only) No
Comments Yes (#) No No
File imports Yes (!import) No No
Query language Yes (TQQ) No jq (external)
Spec maturity 1.0 v3.0 RFC 8259