AI model cost varies by language, token usage can differ by 17 times

A new study reveals that the cost of using AI models varies significantly depending on the language, with token consumption differing by up to 17 times. Published by the AI Language Fairness Research Consortium in September 2026, the research highlights how different languages affect computational costs. Token consumption, a key factor in AI processing, can vary widely even among the most commonly used 20 languages, with some models showing differences of 1.5 to 5 times. When expanded to 200 languages, the gap widens further. Pratish Choudhary, co-founder of Activate AI, notes that non-English languages face challenges in maintaining context and generating responses efficiently. He calls for AI models to treat all languages equally, with transparent cost reporting and targeted improvements. Raul Nanavati of Navana.ai emphasizes the importance of open data projects like RESPIN, which collects text and audio in multiple dialects. However, the study also points out that the use of certain scripts, like Latin letters, can influence token consumption, underscoring the need for better multilingual data and improved tokenizers.