What another language costs
This site exists in Russian and English, which gives a direct comparison: the same pages, two languages.
| Vocabulary | RU, tokens | EN, tokens | RU / EN |
|---|---|---|---|
o200k_base (GPT-6, GPT-5) |
9,332 | 8,072 | 1.16 |
cl100k_base (GPT-4, 2023) |
13,252 | 8,062 | 1.64 |
Measured on 34 of this site’s pages, text only, no code. Claude has its own vocabulary, and for non-English text it
gives more tokens than o200k_base; for the exact number use count_tokens (Tokens). Code, commands and
paths are the same in both languages, so in a prompt with a large config the gap is smaller. Other languages differ
from English by their own ratio — measure on your texts.
What to choose
| Case | Prompt language |
|---|---|
| a chat, an explanation, a task in your own words | the one you think in: precise phrasing matters more than 16% |
| a system prompt sent a million times | English, if money matters; put the answer language in the rules |
| a small local 4–8B model (A model on your own hardware) | English: such models are trained mostly on it and make more mistakes in other languages; how many more, test on your own prompts in both |
| terms, commands, config names | as they are, untranslated: entryPoints, not a translation of it |
Set the answer language explicitly
A model answers in the question’s language, but in a long context full of English documents it may slide into English — or, with documents in another language, out of it. One line in the system prompt settles it: “Answer in German; keep code, commands and terms as in the original.”