There are plenty of token calculators online. All of them assume English. The same content in Korean spends more tokens, so a budget built on those numbers does not hold for a Korean-language service. We worked out how far off it lands, and how that differs by model, across the 225 models with published per-token pricing.
First: what is being compared
Matching character counts makes Korean look 3.65x English, because tokens per character run 0.72 for Korean against 0.20 for English. But saying "a Korean service costs 3.65x" on that basis is wrong. The same content in Korean takes less than half the characters of English. Matching character counts compares different amounts of content.
So we counted the same content in both languages. Five pages of the Python documentation, translated by people, come to 1.31x; the Universal Declaration of Human Rights, 1.14x; 42 paragraph pairs from this site, 1.44x. Across 49,000 tokens that is 1.30x, the figure this piece uses. Samples put it anywhere from 1.1 to 1.4, so read it as an order of magnitude.
What differs is the amount
Say you send 10,000 Korean characters and get 2,000 back, a thousand times a month — summarising or editing a document, around thirty times a day. The 1.30x applies equally to every model, so rankings do not move. What the extra 30% costs, however, does. The median model pays ₩2,082 more a month for Korean; the range runs from ₩60 to ₩624,532, a spread of more than ten thousand times.
| Model | Rate (in/out) | Korean | English | Extra |
|---|---|---|---|---|
| o1-pro | $150 / $600 | ₩2,705,611 | ₩2,081,079 | ₩624,532 |
| gpt-5.4-pro | $30 / $180 | ₩661,371 | ₩508,690 | ₩152,682 |
| kimi-k2.5 | $0.45 / $2.25 | ₩9,019 | ₩6,937 | ₩2,082 |
| mistral-nemo | $0.02 / $0.03 | ₩261 | ₩200 | ₩60 |
o1-pro alone eats an extra ₩625,000 a month for being Korean. mistral-nemo pays ₩60 more for the same reason. Both sit at the same 1.30x.
And it shifts again by model generation
The figures above use o200k, the tokenizer from the GPT-4o generation onward. Counting the same corpus with cl100k, which the generation before it used, Korean comes to 1.16 tokens per character, a factor of 1.6. English does not move at all, at 0.198. The tokenizer change made Korean much cheaper and left English where it was.
So models from an older generation cost more in Korean than a table like the one above says. Anthropic and Google publish no tokenizer, so there is no telling which side they sit on. If you plan to run a lot of Korean, that alone is a reason to pick a current-generation model.
The model matters more than the language
1.30x does not sound like much, and it is not: paying two or three times more because you write in Korean is an overstatement. What matters is where that 30% lands. On an expensive model it is ₩625,000 a month; on a cheap one, ₩60. The choice of model weighs far more than the choice of language.
Look at output too. Korean inflates the reply by the same factor, and output rates run 4.0x input at the median. In the scenario above, 45% of the extra comes from output, rising to 71% on output-heavy models. Picking by the input rate alone is what costs you here.
How to check
The numbers above are one scenario. Different volumes give different conclusions. Paste the prompt you actually send into the calculator and it will line the models up at your own volume. Pasted text never goes into the address, so the result link is safe to share.
Each model page also carries the same scenario, Korean and English side by side.