asia-fertility 🌏

The hidden multilingual tax in your tokenizer — measured before you deploy.

asia-fertility measures the structural penalty that LLM tokenizers impose on lower-resource Asian languages. The same content can cost up to 11.7× more tokens in Burmese than in English on a frontier tokenizer — silent inflation of API bills, smaller usable context windows, and fewer in-context examples.

Quickstart

pip install "asia-fertility[oai,hf,niah,viz]"   # or: uv tool install
source .env                                      # OPENROUTER_API_KEY + HF_TOKEN

# Measure your own text
asia-fertility measure --text "தமிழ் ஒரு செம்மொழி" --lang tam --tokenizer openai/o200k_base

# Compare cost across providers (in your local currency)
asia-fertility cost --text "Xin chào" --lang vie \
  --models openai/gpt-4o,openai/gpt-3.5-turbo \
  --currencies USD,VND

# Reproduce the full leaderboard
asia-fertility run --config configs/study_main.yaml
asia-fertility figures --run runs/main --out runs/main/figures
asia-fertility leaderboard --run runs/main --out runs/main/leaderboard.json

# Multi-turn needle-in-haystack benchmark
asia-fertility niah run --config configs/niah_main.yaml --yes
asia-fertility niah report --output-dir runs/niah/main

# Wall-clock latency benchmark
asia-fertility latency run --config configs/latency_main.yaml --yes
asia-fertility latency report --output-dir runs/latency/main

What's inside

• 16 lower-resource Asian languages: Vietnamese, Indonesian, Malay, Filipino, Thai, Hindi, Bengali, Sinhala, Tamil, Telugu, Kannada, Malayalam, Burmese, Khmer, Lao, plus English baseline.
• 10 tokenizers: OpenAI o200k_base/cl100k_base/o200k_harmony, Mistral Tekken, Qwen3, DeepSeek v3, BLOOM, Gemma-2, Aya Expanse, Llama-3.1. Anthropic Claude and Google Gemini registered as count-only.
• 5 metrics with 95% bootstrap CIs: fertility, premium, same-content cost ratio, characters/token (CPT), and bytes/token (BPT) — the only cross-script-fair comparator.
• Three benchmarks:
  • Leaderboard — same-content cost ratio + BPT across 16 langs × 10 tokenizers on FLORES-200.
  • NIAH multi-turn recall — script-native needle-in-haystack across 5 frontier models × 16 languages × 5 fill levels (4k–131k) × 5 marker positions × 2 trials = 4 000 cells on OpenRouter.
  • Wall-clock latency — streaming-TTFT timing across 5 models × 16 languages × 10 trials (+3 warmup) = 800 measured trials with prefix-cache evasion.

Key findings

• Same content costs 7–12× more tokens on cl100k_base for Brahmic-derived scripts (Tamil 7.61×, Burmese 11.66×). Switching to o200k_base cuts the penalty 3–6×.
• Gemma-2 is the best open-weight tokenizer for South Asian workloads (Tamil 2.58×, Burmese 4.80×). BLOOM dominates Indic scripts (Tamil 1.29×).
• NIAH recall is script-driven, not depth-driven, on 4 of 5 frontier models (gpt-4o-mini, llama-3.1-8b, qwen-2.5-7b, deepseek-v3): non-Latin recall collapses to 0–2/10 already at 4 k context. The exception is gemini-2.5-flash, which holds 76–96% recall across all 16 languages and all 5 fills (4k → 131k) — so the collapse is a vendor-level training choice, not a fundamental limit. Pooled recall across the 4 000 cells: 36.2%. See paper §4.4.
• "Effective context window" is language-dependent: deepseek-v3's nominal 128k window recalls cleanly at 131k on Latin scripts but errors with HTTP 400 on 100% of high-fertility Brahmic-script haystacks at the same notional size (Brahmic fertility blows the prompt past the model's window).
• Cost is NOT a reliable proxy for UX impact: Pearson r between cost ratio and latency ratio = 0.314 pooled across 75 (model, lang) cells; per-model r ranges from −0.19 (qwen) to +0.74 (llama). The relationship is a property of the provider's serving stack (continuous batching, prefill parallelism), not of the language. See paper §4.5.

Paper

Full writeup at paper/paper.pdf. Cite as:

@misc{pedretti2026asianlanguagetax,
  title  = {The Asian Language Tax: Quantifying the Cost, Context, and Recall Penalty of Tokenizing Lower-Resource Asian Languages in Frontier LLMs},
  author = {Pedretti, Antoine},
  year   = {2026},
  url    = {https://github.com/Helmo21/asia-fertility},
}

Data

All three benchmarks are published as a HuggingFace dataset with separate configs:

from datasets import load_dataset

ds   = load_dataset("Helmo21/asia-fertility", "leaderboard")  # cost + BPT leaderboard
niah = load_dataset("Helmo21/asia-fertility", "niah")          # multi-turn recall cells
lat  = load_dataset("Helmo21/asia-fertility", "latency")       # wall-clock timing cells

License

MIT © 2026 Antoine Pedretti. Bundled FLORES-200 data: CC-BY-SA 4.0 (Meta NLLB).