Updated June 2026. Russian, Chinese, Japanese, Korean and Arabic text costs 2–4× more tokens than English on the frontier models. Here's why — and how to cut it.
If you run an LLM application that serves Russian-speaking users, you are almost certainly paying more than you should — and the reason has nothing to do with the prices on the model pages. It's the tokenizer. The same meaning, expressed in Russian, is broken into far more tokens than in English, and you pay per token. We call it the Cyrillic tax. The same effect hits Chinese, Japanese, Korean and Arabic.
Models don't read characters or words — they read tokens, chunks produced by a byte-pair encoder trained on a text corpus. For English, roughly 4 characters make one token. The encoders behind GPT, Claude and Gemini were trained on English-heavy data, so English gets carved into long, efficient tokens. Languages that were underrepresented in training get shattered into much smaller fragments — sometimes one token per character.
Across 200 representative prompts, the same content in Russian cost more tokens than its English equivalent by:
Gemini's gap is smaller because Google trained its tokenizer on a more multilingual corpus from the start. The headline "2.3×" is the rough average across models for Russian prose. For CJK and Arabic the multiplier can be even higher.
The tax compounds in three places at once:
A model advertised at $5/$25 per million tokens effectively behaves like an $11–$13 / $55–$65 model for Russian-heavy output. Run your real prompt through the calculator in both languages and watch the token count jump.
1. Prefer a multilingual-friendly tokenizer. For Russian/CJK-heavy workloads, Gemini 3.5 Flash and DeepSeek V3.2 tokenize noticeably better than the GPT-4 family. The lower multiplier can beat a lower sticker price.
2. Translate to English at the edges. For non-conversational tasks (classification, extraction, summarization), translate input to English, run the model, translate the result back. The translation step is cheap; the savings on a long pipeline are not.
3. Cap and compress output. Output is where the tax hurts most. Ask for terse answers, set max_tokens, and return structured data (JSON) instead of prose where possible.
4. Cache aggressively. A cached Russian system prompt still bills at ~10% — and since the Russian prefix is large, caching saves more in absolute terms than it would in English.
For multilingual products, the model with the lowest headline price is often not the cheapest in practice. Tokenizer efficiency can swing the comparison entirely. Always price your real workload in the actual language you serve — not in English test prompts.
Their tokenizers were trained on English-heavy data, so Russian is split into more, smaller tokens — typically 2–2.6× more than equivalent English. You pay per token, so the bill rises proportionally.
Gemini 3.5 Flash and DeepSeek V3.2 tokenize Cyrillic more efficiently (1.4–1.7× vs 2.1–2.6×), often making them cheaper for Russian workloads despite differing sticker prices.
For batch and non-conversational tasks, usually yes — the cost of a translation step is small compared to running long Russian prompts and outputs at a 2× token multiplier.
Multipliers measured across 200 prompts, June 2026. Use each provider's official tokenizer for exact counts.