The Typographic Tax: How Persian Orthography (Half-Space, Letter and Digit Variants) Changes Token Cost and Representation in Language Model Tokenizers

Purpose. Tokenizers determine what a language costs in language-model applications and how faithfully its writing is represented, and Persian is known to pay a token premium. Less is known about how sensitive that cost is to Persian's orthographic conventions: the zero-width non-joiner (ZWNJ, "half-space"), the Persian versus Arabic forms of two letters, and the script of digits. We quantify both the cross-language premium and the orthographic sensitivity. Design and methods. We evaluated 15 tokenizers (13 general-purpose, among them GPT-4, GPT-4o, Qwen, Mistral, DeepSeek-V3 and XLM-R, and two Persian-specific; 13 distinct vocabularies) on 488 parallel passages in 18 languages (Belebele) and on 4,000 Persian Wikipedia paragraphs. We apply controlled transformations to the same text (ZWNJ replaced by a space or removed, Persian letters replaced by Arabic forms, digit scripts changed) and estimate paired changes in token count with bootstrap intervals, the marginal cost of a ZWNJ, how it is encoded, and whether it survives an encode–decode round trip. Findings. Among the 11 distinct general-purpose vocabularies, Persian text cost a median 2.6 times as many tokens as parallel English (range 1.10 for XLM-R to 5.29 for GPT-2; 3.28 for GPT-4 and 1.52 for GPT-4o), and Persian ranked among the three to six most expensive of 18 languages for eight of the 11 distinct vocabularies. 89.7% of Persian Wikipedia paragraphs contain ZWNJ (12.7 per 1,000 characters). Correct orthography is penalised: replacing ZWNJ with a space lowered token counts for 13 of 15 tokenizers (by 0.04–7.6%; DeepSeek-V3 7.6%), because a ZWNJ costs between 0 and 2.2 tokens (cl100k 1.03; o200k 0.42; DeepSeek-V3 2.23). Substituting Arabic ي/ك for Persian ی/ک changed token counts by −7.7% to +90.5% (+25.7% for ParsBERT-ZWNJ and +90.5% for GPT2-fa, which were trained on normalised Persian), and Persian digits cost 1.9–6.5% more than Latin digits in all general-purpose tokenizers. Three tokenizers (XLM-R, mBERT, ParsBERT-ZWNJ) did not return the ZWNJ after an encode–decode round trip. Originality. The study isolates a within-language orthographic cost that cross-language audits do not measure, shows that the cheapest text is not the best-written text, and gives concrete normalisation guidance for Persian pipelines.