HOMODYNE BENCHMARK LAB / HMD-005
01 / 14 · 00:00

Fewer letters. More tokens.

CHARACTERS · fa ÷ en (median)0.88×

Persian is written shorter.

TOKENS · fa ÷ en (median, 9 tokenizers)1.21–4.61×

…but counted longer.

50 self-written pairs · measured 2026-11-02 · tiktoken 0.9.0 · transformers 4.49.0

HOMODYNE BENCHMARK LAB / HMD-005
02 / 14 · 00:45

What you pay for is not letters. It is tokens.

01What is a token — see it.
02Persian vs English on 9 real tokenizers.
03Which typing habit moves the number.

حدس بزن: فارسی چند برابر انگلیسی؟ ۱٫۵؟ ۲؟ ۳؟

HOMODYNE BENCHMARK LAB / HMD-005
03 / 14 · 01:45

One English sentence = 15 tokens

I want to cancel my subscription and get a refund for the last month.

o200k_base · each box = one token · frequent chunks get one box

HOMODYNE BENCHMARK LAB / HMD-005
04 / 14 · 02:40 · GUESS FIRST

Same sentence in Persian: 22 · 46 · 69

میZWNJ·خواهم اشتراکم را لغو کنم و هزینهZWNJ·ی ماه آخر را پس بگیرم.
English15
o200k_base22
cl100k_base46
gpt2 (r50k)69

gpt2 splits «آ» into two raw bytes: <20d8> <a2>

HOMODYNE BENCHMARK LAB / HMD-005
05 / 14 · 03:35
…differences up to 15 times in some cases.Petrov, La Malfa, Torr, Bibi · Language Model Tokenizers Introduce Unfairness Between Languages · NeurIPS 2023 · arXiv 2305.15425

Tokenizers are trained on text. Not all languages get equal room.

HOMODYNE BENCHMARK LAB / HMD-005
06 / 14 · 04:35
$ python3 artifact/count_tokens.py corpus 50 pairs · chat news howto support code numbers address tiktoken o200k_base · cl100k_base · gpt2 hf Llama-2 · Llama-3.2 · Gemma-3 · Qwen3 · DeepSeek-V3 · Mistral-Nemo ratio fa_tokens ÷ en_tokens (per pair, then median) chars fa ÷ en median = 0.88

HF revisions pinned to exact commits · special tokens not counted

HOMODYNE BENCHMARK LAB / HMD-005
07 / 14 · 05:35

Median Persian ÷ English tokens

Gemma-31.21×
Mistral-Nemo1.29×
Llama-3.21.39×
o200k_base1.39×
DeepSeek-V31.63×
Qwen32.30×
cl100k_base2.92×
Llama-23.42×
gpt2 (r50k)4.61×

measured 2026-11-02 · tiktoken 0.9.0 · transformers 4.49.0

HOMODYNE BENCHMARK LAB / HMD-005
08 / 14 · 06:40

o200k_base: half the pairs fall in 1.23–1.50×

CODE1.20×
ADDRESS1.26×
NUMBERS1.36×
CHAT1.41×
SUPPORT1.44×
NEWS1.45×
HOWTO1.72×

min 1.05 · max 2.10 · Persian higher in 50/50 pairs (o200k) · Gemma-3: 7 pairs where Persian was cheaper

HOMODYNE BENCHMARK LAB / HMD-005
09 / 14 · 07:45 · RE-HOOK

ی vs ي · ک vs ك — same look, different tokens

o200k_base+6.7%

median change, Arabic yeh/kaf

Llama-3.2 · Gemma-3+11.5%

worst sentence: +46.7% / +50%

Mistral-Nemo+10.5%

older tokenizers: ~0 or negative

49 Persian sentences rewritten with U+064A / U+0643 · Persian side only

HOMODYNE BENCHMARK LAB / HMD-005
10 / 14 · 08:50

ZWNJ: free on most · costly on one. Digits: depends.

ZWNJ → SPACE · MOSTmedian ±0%
ZWNJ · DEEPSEEK-V3half-space = 2 byte tokens; ZWNJ → space saves ~13.5%
LATIN DIGITS · o200k / cl100k−8.3% / −21.2%
LATIN DIGITS · GEMMA-3+14.3% (Persian digits cheaper)

فارسی را برای توکن خراب ننویس؛ نیم‌فاصله برای خوانایی و جست‌وجو مهم است.

HOMODYNE BENCHMARK LAB / HMD-005
11 / 14 · 09:55
# same content, same price list input_cost = tokens_in × price_in # price: official page, on the day output_cost = tokens_out × price_out # priced separately cost_fa ≈ ratio × cost_en # 1.39 → +39% · 2.92 → ~3× window_fa ≈ window_en ÷ ratio # 1 ÷ 1.39 ≈ 72% of the content

latency not measured · no current price stated

HOMODYNE BENCHMARK LAB / HMD-005
12 / 14 · 10:55 · CLAIM BOUNDARY
THIS SHOWSOn 50 self-written pairs, Persian took 1.21–4.61× the tokens of English, depending on the tokenizer; spelling choices shift it.
THIS DOES NOT SHOWThat AI understands Persian worse · that every Persian text has this ratio · which tokenizer your product uses today · latency.
HOMODYNE BENCHMARK LAB / HMD-005
13 / 14 · 11:45
1.2–4.6×Persian tokens per English token · 9 tokenizers · 50 pairs

Your text. Your tokenizer. Count it → python3 artifact/count_tokens.py

پس: tokenizerها تکه‌های فارسی کمتری می‌شناسند و متن ما را ریزتر می‌شکنند.

HOMODYNE BENCHMARK LAB / HMD-005
14 / 14 · 12:30 · NEXT

Next: “fewer errors” — out of what?

MODEL A26% errors
MODEL B75% errors
THE OTHER COLUMN?

SimpleQA table · OpenAI, Sept 2025 (R01) · HMD-006