Why Talking to an AI in Russian Costs More and How to Work With It

When I build agents and bots on n8n and the API for clients, the token bill is almost always above the forecast if the whole dialogue runs in Russian. It's not the provider's greed. It's how the model cuts text into tokens before processing it.

What a token is and why it's not a letter or a word

An AI doesn't read text letter by letter. It cuts it into pieces, tokens, and turns those into numbers for calculation. For English, tokenizers are trained well: one token often covers a whole word or syllable, on average about 4 to 5 characters per token.

With Cyrillic it doesn't work out that way. For most tokenizers, Russian text comes to roughly one token per letter. The same thought that fits in one token in English may need three or four in Russian.

The difference in numbers, not in words

Independent testing of tokenizers (the methodology is described on dev.to, where the author checked through tiktoken and model APIs, relying on earlier work by Petrov et al., NeurIPS 2023) gave concrete multipliers for the same meaning in two languages:

One technical text from the same testing took 327 tokens in English and 872 tokens in Russian on Claude Opus 5. That is 2.67 times for the very same meaning, without a single extra word.

Why it turned out this way

Two reasons add up. The first is technical: Cyrillic in UTF-8 is encoded with two bytes per character, Latin with one, and the tokenizer feels it. The second is about training: the tokenizer's vocabulary is built on a corpus of texts, and the corpora of the large models are skewed toward English. The model effectively "knows" English words whole, while Russian ones it has to assemble from pieces.

What this changes in practice

If you work in a ChatGPT or Claude chat and just exchange messages, the difference in money isn't noticeable. It starts to matter where Russian text passes through the API many times in a row.

The main case: an agent or bot that processes incoming messages through the API and pays for every token. With the same volume of conversation, a Russian-language bot costs one and a half to three times more per month than exactly the same one in English. A similar story with a long conversation in one window: a model's context window is limited by the number of tokens, not words, and a Russian dialogue fills it faster, so the model starts to "forget" the beginning of the conversation sooner.

Separately there are system instructions and templates that the model reads on every request. If the instruction for an agent is written in Russian and loaded into every call, you pay the same tax on Cyrillic many times over, for the same text.

How to cut the cost without losing quality

In the agents I build for clients, I usually write the system instructions and data schemas in English, and the model still phrases the final reply to the client in Russian. The model reads the instruction on every call, and the token savings on this part are really noticeable in the monthly bill.

One more point: don't send the model the whole previous text again if you need to fix one paragraph. Ask for a targeted edit, not a full regeneration, which also noticeably lowers input token use. And if you're working with a long document in Russian, ask the model to first compress it into key points, and then work with the key points, not the full text in every new message.

Bottom line

The difference between Russian and English in tokens is not a trifle and not a myth from a Telegram channel: it has a published methodology and concrete numbers by model. For a one-off question in a chat it isn't critical. For a bot, an agent, or any system that counts tokens automatically, it is a direct expense line that you can and should work with.

I write about how I build such systems for experts and what costs are really built into them in the channel «Растём с ИИ» ("Growing with AI"). If you're counting the budget of your bot or agent and want to go through it soberly, sign up for a free consultation.