Tashkent, Uzbekistan (UzDaily.uz) — The same text written in Uzbek Latin script requires 1.55–2.17 times as many tokens as its English equivalent across five major AI providers, making it more expensive to process. This is the finding of a study by Gorgona AI titled “One Meaning, Different Prices: How Much Does Uzbek Cost AI?”
The researchers compared OpenAI, Anthropic, Google, Qwen and DeepSeek. They used 1,337 parallel texts, meaning translations of the same content, and processed them through 15 tokenisers. The results were then checked against actual model responses.
Language models do not work with letters or words directly, but with tokens. These are fragments into which a tokeniser divides text. Its vocabulary is built from training data, which contains a much larger share of English. As a result, an English word often takes up a single token, while an Uzbek word with its suffixes is split into almost three tokens on average.
Providers charge the same rate per token regardless of language. Context window capacity is also measured in tokens, and models generate responses one token at a time. If the same meaning in Uzbek requires twice as many tokens, the bill doubles. The context window can hold half as many documents, while generating a response at the same speed takes twice as long.
The additional cost varies by model and text. Texts originally written in Uzbek tend to be closer to the lower end of the range, while translations from English are closer to the upper end. In the FLORES dataset, which includes news articles, encyclopaedia entries and travel guides, the difference ranged from 1.78 times for OpenAI to 2.13 times for DeepSeek.
Russian texts from the same dataset cost 1.35–1.60 times more to process than English. Uzbek is 1.24–1.54 times more expensive than Russian. According to the researchers, genre has little effect on these ratios: the difference between genres for each model does not exceed 8%.
Morphology alone does not explain the problem. The study cites a 2026 analysis that tested 30 identical sentences using the GPT-4o tokeniser. The English text required 348 tokens, Russian 551, Turkish 609, Uzbek in Latin script 763 and Uzbek in Cyrillic 814.
Turkish is also an agglutinative language written in the Latin alphabet, but it required one-fifth fewer tokens. Turkish was cheaper to process than Uzbek with 13 of the 15 tokenisers. The exceptions were GigaChat 3 and NeuronAI-Uzbek.
The researchers argue that the additional cost is determined not by the number of native speakers but by how much Uzbek text models have encountered during training and how much space the language occupies in a tokeniser's vocabulary.
The choice of model has the greatest impact on the final cost. Claude Opus 5.5 and GPT-5.6 Sol have the same listed price per token. However, processing an Uzbek translation of a text that occupies one million words in English costs US$14.78 with Claude and US$8.70 with GPT, a difference of 1.7 times.
Claude's tokeniser uses 1.57 times as many tokens as OpenAI's for English text and 1.70 times as many for Uzbek in Latin script.
The study uses official prices as of 14 October 2026, excluding caching. The cost of processing the same Uzbek text ranges from US$0.37 with Claude Haiku 5.5 to US$14.78 with Claude Opus 5.5.
The authors therefore recommend comparing models based on the cost of processing a particular text rather than the price per million tokens. They emphasise, however, that their findings concern costs only; they did not measure the quality of model responses.
Responses in Uzbek are also more expensive to generate than those in English. GPT-5.6 Luna and Claude Haiku 5.5 were each given 3,904 prompts to summarise a paragraph and provide a detailed answer to a question about it in four language variants.
The visible Uzbek responses cost 1.7–1.8 times more than the English ones, although their character counts were almost identical.
Claude Haiku 5.5, without special settings, decided independently whether to perform reasoning before responding. It did so more frequently in Uzbek: in 28% of summaries, compared with 10% in English, and in 95% of answers to questions, compared with 63%.
Users cannot see this reasoning, but they are charged for it as output tokens. When reasoning is included, a paid response in Uzbek costs 2.1–2.3 times more than an English one. The researchers caution that this observation is based on a single model and two tasks.
The cost difference between Uzbek Cyrillic and Latin scripts also depends on the model. Claude charges the same for both alphabets. With other providers, Cyrillic is 7–23% more expensive, with the largest difference recorded for DeepSeek.
The researchers attribute this not to the structure of the Cyrillic alphabet but to the amount of text in each script that models encountered during training.
The apostrophe used in the Uzbek letters oʻ and gʻ also adds to the cost. It is currently typed using four different characters, which tokenisers treat as distinct symbols.
In the NTREX news dataset, the original spelling was 0.6% more expensive than the same text using a straight apostrophe with OpenAI, 1.9% with Claude and 4.2% with Qwen. Gemini normalises apostrophes into a single form and therefore showed no difference.
The researchers separately assessed the potential impact of the transition to a new alphabet. The law introducing it was signed on 21 October 2026 and will take effect on 22 April 2027. The letters oʻ and gʻ will be replaced by ö and ğ, while sh and ch will become ş and ç. The combination ng will no longer be treated as a separate letter.
Under the new spelling system, texts become 4% shorter in terms of character count. The number of tokens falls by 2–4% with Claude, Gemini and Qwen, while changes with OpenAI and DeepSeek are negligible.
The study therefore concludes that companies will not need to revise their AI budgets because of the alphabet change. The authors see the main benefit in giving each word a single spelling. This will make document searches, duplicate removal and the use of AI assistants with knowledge bases easier.
Independent researcher Dusmamat Karimov reached similar conclusions about the alphabet. In October, he tested around 460,000 characters from Uzbek Wikipedia and Gazeta.uz news articles using seven tokenisers.
Open models show that the vocabulary used by a tokeniser determines the additional cost. The same Uzbek text requires 25% fewer tokens with GigaChat 3 than with OpenAI's tokeniser. With NeuronAI-Uzbek, the difference is only 2%.
YandexGPT processes Russian efficiently, but Uzbek costs twice as much to process as Russian.
Over time, Uzbek has become cheaper more slowly than major languages. When OpenAI switched from the GPT-4 tokeniser to GPT-4o, the additional cost for Uzbek texts in the FLORES dataset fell from 2.19 to 1.78, a reduction of around one-fifth.
With Qwen, the transition from the third version to 3.5 reduced the cost of Uzbek in Latin script by only 6%, while the cost for Russian fell by 23%.
The researchers regard open Uzbek-language text collections as digital infrastructure: the more data available, the cheaper AI services in Uzbek will become for everyone, from start-ups to government agencies.
By April 2027, they propose preparing parallel text collections in both alphabets, open conversion tools and standardised typing rules.
According to the study, creating Uzbek-language text collections, morphological dictionaries, tokenisers and quality tests will require cooperation between companies, universities and the state, as well as dialogue with model developers.
