Tokens in Japanese

What the numbers say

The same website copy is 200 tokens in English and 313 tokens in Japanese on the newer o200k_base encoding, a ratio of 1.56×. On the older cl100k_base encoding the Japanese copy needs 421 tokens, which is 2.10× the English count of 200. At the same price per token, the same message costs about 1.56× as much in Japanese on the newer encoding. Among the 29 non-English languages we measured, Japanese ranks number 27 from the cheapest.

The newer encoding made a real difference: o200k_base used 26% fewer tokens than cl100k_base for Japanese, compared with 34% across all the non-English languages in the data. That is less than average, so Japanese gained less from the newer vocabulary than most.

Why Japanese? Japanese mixes three scripts in one sentence: kanji (borrowed Chinese characters), and the hiragana and katakana syllabaries. Particles and verb endings in hiragana are very common and very short, which means a sentence has many small pieces for a tokenizer to handle, and rarer kanji can be split into several byte-level tokens.

East Asian scripts have thousands of characters, each of which is three bytes in UTF-8. A tokenizer cannot include every character as its own piece, so common characters get a single token and rarer ones are broken into byte-level pieces. At the same time these languages need far fewer characters to say the same thing, which is why total cost, not characters per token, is the number to compare.

What it means for your bill

API prices are per token, so the ratio is also a price ratio. If a prompt costs $1.00 in English on o200k_base, the equivalent Japanese prompt costs about $1.56, and about $2.10 on cl100k_base. A document of 1,000 Japanese characters is roughly 660 tokens on o200k_base. These are estimates from one sample: translated text is never exactly the same length as the original, and your text may be formal, technical or full of names, which changes the ratio.

To turn tokens into money, enter your provider's input and output prices per million tokens in the token calculator. For how pricing works, see how LLM API pricing works.

How to use fewer tokens with Japanese text

1. Use a model with the newer tokenizer. For Japanese, o200k_base used 26% fewer tokens than cl100k_base in this sample, so a GPT-4o-class model needs fewer tokens than an older one for the same text.
2. Count before you send. Paste a typical message into the token calculator to see the real count, then multiply by the number of requests.
3. Keep fixed instructions short. A system prompt that is sent with every request is paid for every time. Writing the standing instructions in English is often cheaper per token than in Japanese; test that answers stay as good as you need.
4. Cut what is not needed. Remove greetings, repeated context and long examples. The guide to reducing token usage has more.
5. Be careful with emoji, rare characters and unusual formatting. They are often split into several byte-level tokens each.