Your AI costs more in Hindi than in English — token cost across 14 languages

If you run an AI feature in more than one language, you are not paying the same price in each one. The same sentence costs a different number of tokens depending on the language it is written in.

We measured it across 14 languages. The short version: the writing system does not predict the cost. German costs more than Korean. Russian costs more than Japanese. Chinese is almost as cheap as English.

How we measured

• Encoding o200k_base via tiktoken 0.14.0 — measured 13 October 2026
• One sentence with the same meaning in 14 languages: “Please summarize this document in three sentences and list the main points.”
• English is the baseline (13 tokens)
• Input tokens only

The results

Three things that surprised us

1. Chinese is nearly as cheap as English (1.08x). People assume non-Latin scripts are expensive. Chinese characters are dense — the whole sentence is 18 characters — and the tokenizer handles them well. It came second cheapest out of 14.

2. German (1.46x) costs more than Korean (1.38x). A Latin-script European language is more expensive than a language with its own alphabet. German’s long compound words get split into several pieces each.

3. Russian (1.85x) costs more than Japanese (1.54x) and Korean (1.38x). Cyrillic is an alphabet much like Latin, yet it is one of the most expensive scripts here. Script family tells you very little.

Latin and CJK come out almost identical on average (1.32x vs 1.33x), which is the opposite of what most people expect.

What this costs at scale

Per 100,000 requests, for this one sentence:

That is one short sentence. A real prompt carries a system prompt, instructions and user content — the gap scales with all of it.

What to do with this

• Do not budget one number for all languages. If you price a feature on English usage and then launch in Hindi or Russian, your input cost roughly doubles.
• Measure the languages you actually ship in — not a rule of thumb about scripts, because the script does not predict the cost.
• Do not translate to English to save money. The translation call costs tokens too, and you lose nuance. The savings here are small compared to that risk.
• Where it matters most is repeated text — a system prompt sent on every request multiplies this difference by your request count.

Try it yourself

import tiktoken
enc = tiktoken.get_encoding("o200k_base")
n = lambda t: len(enc.encode(t))

sentences = {
    "English": "Please summarize this document in three sentences "
               "and list the main points.",
    "German":  "Bitte fassen Sie dieses Dokument in drei Satzen zusammen "
               "und listen Sie die wichtigsten Punkte auf.",
    # ... add the languages you actually ship in
}

base = n(sentences["English"])
for lang, s in sentences.items():
    print("%-12s %3d chars  %3d tokens  %.2fx English"
          % (lang, len(s), n(s), n(s) / base))

Limits of this measurement

• We wrote the translations ourselves. They are accurate in meaning but may not be the most idiomatic phrasing a native speaker would choose, and phrasing changes token counts. Treat the ordering as indicative, not exact.
• This is one sentence. A single short sentence cannot represent a whole language. Longer text, technical vocabulary and names would all shift the numbers.
• One tokenizer only (o200k_base). Models using a different tokenizer will split these languages differently — sometimes very differently.
• Input tokens only. Output is often priced higher, and output length also varies by language, which we did not measure.
• We measured cost, not quality. How well a model answers in each language is a separate question and is not addressed here.

Measured 13 October 2026 · tiktoken 0.14.0 · o200k_base · input tokens

Read next

• AI Terms Explained: A Practical Glossary for Reading Tech News
• How to Prompt AI Assistants: Five Principles That Actually Work
• Seven Rules for Using AI at Work Without Creating Problems

About the Author

MAILY POST Editorial

Administrator

MAILY POST publishes practical guides for people actually using AI. Rather than reporting what was announced this week, we answer the question that comes after: how do you actually use this, and what should you watch out for. Guides are labelled as guides — they are compiled from documentation and common practice, not from experiments we ran. When we do measure something ourselves, we publish the conditions, the code to reproduce it, and the limits of what we found. We correct errors when readers report them through our contact page.

Visit Website View All Posts