Listen to the article
The same text written in Uzbek Latin script requires between 1.55 and 2.17 times as many tokens as its English equivalent across the five largest AI providers, making Uzbek-language applications measurably more expensive to run, according to a study by Gorgona AI published under the title One Meaning, Different Price: How Much Does the Uzbek Language Cost for AI?
The researchers compared tokenisers from OpenAI, Anthropic, Google, Qwen and DeepSeek, running 1,337 parallel texts — identical content translated across languages — through 15 tokenisers and then verifying results against live model outputs. Because AI language models bill by the token, charge the same per-token rate regardless of language, and measure context windows in tokens, a text that requires twice as many tokens doubles the invoice, halves the number of documents that fit in a single context window, and takes twice as long to generate at equivalent speed.
Why Uzbek costs more
Tokenisers build their vocabularies from training data, which skews heavily toward English. An English word typically occupies a single token; an Uzbek word, with its agglutinative suffixes, fragments into nearly three on average. The study notes that the penalty is not purely morphological: Turkish is also agglutinative and also uses Latin script, yet it required roughly one-fifth fewer tokens than Uzbek Latin on the GPT-4o tokeniser in a 2026 study cited by the authors, where 30 identical sentences produced 348 tokens in English, 551 in Russian, 609 in Turkish, 763 in Uzbek Latin and 814 in Uzbek Cyrillic. Turkish was cheaper than Uzbek on 13 of the 15 tokenisers tested. The authors attribute the difference to the volume of Uzbek text seen during model training and the share of vocabulary slots allocated to Uzbek, not to the structure of the alphabet itself.
On the FLORES corpus — covering news, encyclopaedia articles and travel guides — the markup over English ranged from 1.78 times at OpenAI to 2.13 times at DeepSeek. Russian on the same texts costs 1.35 to 1.60 times more than English; Uzbek costs a further 1.24 to 1.54 times more than Russian. The authors found that genre had little effect on these ratios: within any single model the variation across text types was no more than eight percent.
Model choice matters more than language alone
The study illustrates how provider selection amplifies the language penalty. Claude Opus 5.5 and GPT-5.6 Sol carry identical list prices, yet a one-million-word English text translated into Uzbek would cost $14.78 to process as input on Claude against $8.70 on GPT — a 1.7-fold difference — because Anthropic’s tokeniser already uses 1.57 times more tokens than OpenAI’s on English text, rising to 1.70 times more on Uzbek Latin. All prices in the study are drawn from official provider rate cards as of 14 October 2026, excluding caching. Across the full range of models tested, the cost of the same Uzbek input varied from $0.37 on Claude Haiku 5.5 to $14.78 on Claude Opus 5.5, which the authors say makes per-million-token pricing a poor basis for budget comparisons; they recommend benchmarking against one’s own actual texts instead.
Output costs follow the same pattern. In a test of 3,904 prompts sent to both GPT-5.6 Luna and Claude Haiku 5.5 — asking each model to summarise a paragraph and answer a question about it in four language variants — visible Uzbek responses cost 1.7 to 1.8 times more than English responses despite producing nearly the same number of characters.
A separate finding concerns hidden reasoning tokens. Claude Haiku 5.5, when given no explicit instructions, chose to reason internally before answering. It did so in 28 percent of Uzbek summarisation tasks versus 10 percent in English, and in 95 percent of Uzbek question-answering tasks versus 63 percent in English. Users do not see this reasoning but are billed for it as output tokens. Once reasoning was included, the effective cost of a Uzbek response reached 2.1 to 2.3 times the English equivalent. The authors caution that this observation covers only one model and two task types.
Cyrillic, apostrophes and the coming alphabet reform
Among providers other than Anthropic, Uzbek Cyrillic costs 7 to 23 percent more than Uzbek Latin, with the largest gap at DeepSeek. Claude treats both scripts identically. The authors attribute the Cyrillic premium to training-data volume rather than script complexity.
Even the apostrophe variant used in the letters oʻ and gʻ carries a measurable cost. Four different Unicode characters are currently used interchangeably for this apostrophe, and tokenisers treat them as distinct symbols. On the NTREX news corpus, texts using mixed apostrophe forms cost 0.6 percent more than texts standardised to a straight apostrophe at OpenAI, 1.9 percent more at Anthropic and 4.2 percent more at Qwen. Google’s Gemini normalises apostrophes automatically and recorded no difference.
Uzbekistan’s new alphabet law, signed on 21 October 2026 and taking effect on 22 April 2027, replaces oʻ and gʻ with ö and ğ, substitutes ş and ç for sh and ch, and removes ng as a digraph letter. The reformed script shortens texts by about four percent in character count. Token counts fall by two to four percent on Claude, Gemini and Qwen tokenisers but change negligibly on OpenAI and DeepSeek. The study concludes that the reform will not require companies to revise AI budgets materially; its main practical benefit, the authors argue, is that each word will have a single canonical spelling, simplifying document search, deduplication and knowledge-base retrieval.
Open data as infrastructure
The study’s findings align with independent work by Dusmamat Karimov, who in October tested roughly 460,000 characters from Uzbek Wikipedia and gazeta.uz across seven tokenisers. Open-source models demonstrate that the premium is driven by vocabulary design: GigaChat 3 processes the same Uzbek text with 25 percent fewer tokens than OpenAI’s tokeniser; NeuronAI-Uzbek reduces the gap to just two percent. YandexGPT handles Russian efficiently but charges twice as much for Uzbek as for Russian.
Progress has been slow. When OpenAI moved from the GPT-4 to the GPT-4o tokeniser, the Uzbek markup on FLORES texts fell from 2.19 to 1.78 — a reduction of roughly one-fifth. At Qwen, the transition from version 3 to 3.5 cut the Uzbek Latin premium by only six percent, while Russian became 23 percent cheaper.
The authors frame open Uzbek text corpora as digital infrastructure, arguing that expanding them would reduce AI costs for everyone from startups to government agencies. Ahead of the April 2027 alphabet transition, they recommend preparing parallel corpora in both scripts, open-source conversion tools and unified typographic standards. Building the full stack — corpora, morphological dictionaries, tokenisers and quality benchmarks — will require coordinated effort from companies, universities and the state, working in dialogue with model developers, the study concludes.