The reason is tokens, the units of text or data that AI models process, and the tokenizer, the tool that breaks raw text, images, or audio into those units.
Researchers call the resulting extra cost a hidden "tokenizer tax." Among the 20 most popular languages online, the most expensive used between 1.5 and 5 times more tokens than the cheapest.
Across 200 languages, the gap in one model grew by 17 times.
More tokens mean higher latency (slower responses) and less effective context capacity, so less room for documents or conversation history.
The authors say these penalties hit hardest where computing power and connectivity are already limited, and that tokenizer inequity constrains who can participate in the AI economy.
They offer an example: a startup in Bengaluru may pay a surcharge on every prompt and response that a competitor in Austin does not.
Pratyush Choudhury, co-founder of the AI-focused venture fund Activate AI, said a tutoring app that has to retain a student's earlier mistakes and explanations would use up its context window faster in one language than another, which could reduce the quality of the service.
Language rarity is not the only factor; script matters too. The study says Asturian, which has few speakers, benefits from using the Latin script and from related Romance languages such as Italian. Burmese, Tibetan, Tifinagh, Khmer, Lao, and Ol Chiki, written in scripts tokenizers rarely saw, were consistently the most disadvantaged across nearly all models analyzed. Dheemanth Reddy of Maya Research attributes the problem to a lack of high-quality digital data, but warns that better data alone won't fix the tokenizer. Choudhury expects falling token prices to ease the direct cost over time, yet says labs will need to design future tokenizers deliberately for these languages and scripts. He also urged major labs to publish language-by-language scorecards showing what comparable tasks cost and to commit to measurable improvements for their worst-served languages.