artok عرتوك

Arabic Token Tax Calculator

See how much more Arabic costs across 18 LLM tokenizers.

Website · GitHub · Report Bug

---

Why This Exists

Arabic text uses 2-5x more tokens than equivalent English depending on the tokenizer. Same meaning, wildly different cost:

Text	Claude Tokens	GPT-4.1 Tokens
الذكاء الاصطناعي يغير العالم	25	10
AI is changing the world	5	5
Ratio	5.0x	2.0x

This is the Arabic Token Tax. If you're building Arabic AI products, you're paying 2-5x more for the same capabilities. artok makes this visible, measurable, and actionable.

Install

git clone https://github.com/Moshe-ship/artok.git
cd artok
python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[all]"

Quick Start

# See the tax across all 18 tokenizers
artok "الذكاء الاصطناعي يغير العالم"

# Compare Arabic vs English side by side
artok "الذكاء الاصطناعي" -e "Artificial intelligence"

# Run the Arabic friendliness benchmark (no input needed)
artok --benchmark

# See how diacritics inflate tokens
artok "بِسْمِ اللَّهِ الرَّحْمَنِ الرَّحِيمِ" --tashkeel

# Compare dialects (MSA vs Egyptian vs Gulf vs Levantine vs Moroccan)
artok --dialects

# Heatmap — color each word by token cost
artok "الذكاء الاصطناعي يغير حياتنا" --heatmap

# Rank tokenizers by composite score
artok "نص عربي" --leaderboard

# Estimate monthly costs at 50M tokens
artok "نص عربي" -e "Arabic text" --cost 50

# See savings from switching away from Claude
artok "الذكاء الاصطناعي يغير العالم" --switch-from claude-sonnet

# Compare Arabic against other languages
artok "الذكاء الاصطناعي" --compare-langs 'en:AI|fr:IA|zh:人工智能'

# Analyze Arabic text from a URL
artok --url https://ar.wikipedia.org/wiki/ذكاء_اصطناعي -t gpt4.1,claude-sonnet

# Export as SVG for sharing on X/Twitter
artok "نص عربي" -e "Arabic text" --export results.svg

# Live mode — type and see counts in real-time
artok --watch

# Keep pricing up to date
artok --update

Supported Tokenizers (18)

18 tokenizers across 10 providers. Pricing auto-updates from GitHub (artok --update).

Tokenizer	Provider	Input $/1M	Output $/1M
GPT-4.1	OpenAI	$2.00	$8.00
GPT-4.1 mini	OpenAI	$0.40	$1.60
GPT-4.1 nano	OpenAI	$0.10	$0.40
GPT-4o	OpenAI	$2.50	$10.00
GPT-4o mini	OpenAI	$0.15	$0.60
Claude Opus 4.6	Anthropic	$5.00	$25.00
Claude Sonnet 4.6	Anthropic	$3.00	$15.00
Claude Haiku 4.5	Anthropic	$1.00	$5.00
Llama 4	Meta	$0.18	$0.18
Qwen 3.5	Alibaba	$0.10	$0.40
Mistral Large 3	Mistral	$0.50	$1.50
Mistral Small	Mistral	$0.10	$0.30
Gemini 2.5 Pro	Google	$1.25	$10.00
Gemini 3 Flash	Google	$0.50	$3.00
DeepSeek V3.2	DeepSeek	$0.27	$1.10
Grok 2	xAI	$2.00	$10.00
Command R+	Cohere	$2.50	$10.00
Jamba 1.5	AI21	$0.20	$0.40

All Features

Flag	What it does
(text)	Token count across all 18 tokenizers
-e	Arabic vs English comparison
-t	Filter to specific tokenizers
-c	Cost estimate at N million tokens
-w	Cost estimate at N million words
-f	Read text from file
--json	JSON output for scripting
--chart	Visual bar chart
--viz	Colored token split visualization
--batch	Process JSONL/CSV files
--recommend	Best tokenizer for a budget
--switch-from	Savings from switching providers
--compare-langs	Arabic vs other languages
--url	Analyze Arabic text from a URL
--tashkeel	Diacritics inflation analysis
--heatmap	Color words by token cost
--benchmark	Arabic friendliness score 0-100
--dialects	MSA vs Egyptian vs Gulf vs Levantine vs Moroccan
--leaderboard	Composite score ranking
--watch	Live interactive mode
--export	Export to SVG
--update	Fetch latest pricing from GitHub
--list	Show all tokenizers with source info

The Arabic Token Tax

Most LLM tokenizers are trained primarily on English/Latin text. Arabic characters get split into individual bytes or small fragments instead of whole words. The result: same meaning, 2-5x more tokens, 2-5x higher cost.

Benchmark results (artok --benchmark):

Rank	Tokenizer	Arabic Friendliness Score
1	Mistral Large 3	92.1/100
2	Qwen 3.5	91.7/100
3	GPT-4.1	91.1/100
4	Gemini 2.5 Pro	90.2/100
5	Llama 4	83.6/100
...	Grok 2	50.3/100
...	Claude Sonnet 4.6	25.6/100

Key findings:

• Mistral / Qwen 3.5: Best for Arabic — dedicated multilingual tokenizers
• GPT-4.1 / Gemini / DeepSeek / Llama 4: Good at ~1.5-2.5x vs English
• Grok 2: Moderate at ~2.5-3.5x vs English
• Claude: Worst for Arabic at ~3.5-5x vs English — byte-level encoding for Arabic

Auto-Update Pricing

Tokenizer pricing changes frequently. artok handles this automatically:

1. On startup, checks ~/.cache/artok/tokenizers.json (24h cache)
2. If stale, fetches latest from tokenizers.json on GitHub
3. Falls back to hardcoded defaults if offline

artok --update    # Force refresh right now
artok --list      # Shows pricing source (cached vs built-in)

To update pricing for all users: Edit tokenizers.json in this repo. All users get the new prices within 24 hours.

Architecture

artok/
├── cli.py          # CLI entry point, argument parsing
├── core.py         # Tokenizer loading, counting, remote config
├── display.py      # Rich terminal output (tables, charts, heatmaps)
├── __init__.py     # Version
└── __main__.py     # python -m artok support
tokenizers.json     # Remote-updatable pricing config
docs/index.html     # GitHub Pages landing page

• Tokenizer backends: tiktoken (OpenAI), transformers (HuggingFace), tokenizer_fast (direct tokenizer.json loading for DeepSeek)
• Output: Rich tables, colored text, bar charts, SVG export
• Config: Hardcoded defaults + GitHub-hosted JSON with 24h local cache

Contributing

1. Fork the repo
2. Add a tokenizer to tokenizers.json (and core.py TOKENIZERS list for the hardcoded fallback)
3. Test with artok --list and artok "مرحبا" -t your-new-tokenizer
4. Open a PR

To update pricing only: edit tokenizers.json — no code changes needed.

License

MIT — Musa the Carpenter