Uzbek language for AI costs 1.55–2.17 times more than English
A study by Gorgona AI found that identical text in the Uzbek Latin script requires more tokens from major AI providers, making its processing more expensive.
The authors compared OpenAI, Anthropic, Google, Qwen, and DeepSeek using 1,337 parallel texts and verified the results across 15 tokenizers. Tokens are text fragments that models work with. Providers charge based on their quantity regardless of the language, so more 'expensive' tokenization increases the costs of requests and responses.
On the FLORES corpus, which includes news, encyclopedic articles, and travel guides, Uzbek text required between 1.78 times more tokens for OpenAI and 2.13 times more for DeepSeek compared to English. For the Russian language, the excess ranged from 1.35 to 1.60 times, while Uzbek proved to be 1.24–1.54 times more expensive than Russian. Translations from English typically fell closer to the upper end of the range than texts originally written in Uzbek.
The difference affects not only the cost of requests. Fewer documents fit into the context window, and at the same speed of generation, responses may take longer to produce. The final amount also depends on the chosen model: the cost of Uzbek text in the study varied from $0.37 for Claude Haiku 5.5 to $14.78 for Claude Opus 5.5. The authors emphasize that they compared costs specifically, and did not measure the quality of responses.
According to Gorgona AI’s assessment, the key factors are the volume of Uzbek texts in training data and the language’s position in the tokenizer’s vocabulary. Therefore, the researchers suggest expanding open corpora in Uzbek, creating morphological dictionaries, and developing quality tests in collaboration with companies, universities, the state, and model developers.
