The Token Tax on Indian Languages: Measuring Tokenizer Premiums Across All 22 Scheduled Languages and 14 Tokenizers

Large language models (LLMs) process text as subword tokens, and the number of tokens a text requires determines API cost, usable context length and generation latency. Tokenizers built from English-dominated data segment other languages into more tokens, imposing a "token tax". We measure this tax for all 22 scheduled languages of India using 1,024 professionally translated parallel sentences (IN22-Gen) and 14 tokenizers, including GPT-4o, Llama 4, Gemma 3, DeepSeek-V3, Qwen 3 and the Indian-built Sarvam-1, Krutrim-2, IndicBERT v2 and MuRIL. Relative to English, GPT-4o needs 3.10x as many tokens on average (1.55x for Hindi, 13.2x for Santali), and the most efficient LLM tokenizer, Gemma 3, needs 2.33x. Newer tokenizers roughly halve the premium of their predecessors, but not for the scripts with the least web presence. Sarvam-1 is the most efficient LLM tokenizer on its ten target languages (1.28x) yet less efficient than GPT-4o on the remaining twelve. Using human romanizations from the Dakshina dataset, we find that romanized input is cheaper only for older, English-centric tokenizers and costs 38-44% more on Indic-focused ones. Finally, Unicode NFC normalization changes token counts by between -17.4% and +7.6% depending on the tokenizer. We also document a library-level discrepancy that inflated GPT-4o counts by about 46% and describe the verification protocol that caught it. Code and data are released.