How many more tokens does Swedish need than English?
Swedish needs 35 % more tokens than English for the same text on ChatGPT / Copilot. So your quota runs out about 35 % faster, chat memory fills up just as much faster, and answers take just as much longer. Swedish ranks 192 of 203 languages, where 1 is the most expensive. Across 6 current AI tokenizers, the token tax ranges from 32 % (Claude) to 55 % (DeepSeek).
| Model | Tokenizer | Token tax |
|---|
| ChatGPT / Copilot | o200k_base | 35 % |
| Gemini | Gemini 3.x (unpublished; = Gemma 4 on text) | 36 % |
| Claude | Claude (unpublished) | 32 % |
| Meta AI / Llama | Llama 4 | 32 % |
| DeepSeek | DeepSeek V3 | 55 % |
| Qwen | Qwen 3.5 (248k) | 37 % |
What is the token tax?
The token tax is tokens for a language ÷ tokens for English on the same text per tokenizer. Petrov et al. (2023) introduced this concept to make tokenization easier to compare across languages, but call it the tokenization premium. There is no established term for it yet; it is also called the tokenization premium or the language tax.
Tokenizers learn their vocabulary mostly from English text, so common English words fit in one token while Swedish text is split into more, shorter pieces. Languages with little training text and non-Latin scripts tend to pay the most.
Petrov et al. sum up the problem: “The same text translated into different languages can have drastically different tokenization lengths” (NeurIPS 2023).
How was this measured?
Each tokenizer counted the tokens in the same 1,012 sentences of the FLORES-200 devtest set, in Swedish and in English, following Petrov et al. (NeurIPS 2023). Counted 2026-10-26.
What is a token?
AI models read text as tokens. Before a model sees your text, its tokenizer cuts it into pieces. You pay per token, chat memory is counted in tokens, and answers are written one token at a time. Here is a science news headline from October 2026 as ChatGPT / Copilot cuts it, in Swedish and in English:
En enda fiollektion förändrar hjärnvågsmönstren när nybörjare lär sig spela
En | enda | fi | ol | lektion | föränd | rar | hj | är | nv | åg | sm | ön | str | en | när | ny | bör | j | are | lär | sig | spela = 23 tokens
One violin lesson changes brain wave patterns as beginners learn to play
One | violin | lesson | changes | brain | wave | patterns | as | beginners | learn | to | play = 12 tokens
For this one sentence the ratio is 23 ÷ 12 = 1.92: 92 % more tokens than English. One sentence is too little text for a reliable number, which is why the numbers on this page are measured on 1,012 FLORES-200 sentences.
How does Swedish compare with other languages?
Swedish pays 35 % on ChatGPT / Copilot, while the median of all 203 languages pays 75 %.
How can I reduce the token tax?
The simplest way is to chat in English, if you are comfortable with that. Other ways are hard to recommend, because models differ not only in token tax but also in how good an answer you get per token.
Questions about the token tax
Does every language pay a token tax?
Almost every one. On ChatGPT / Copilot, all 203 languages need more tokens than English, from 23 % (Portuguese) to 1 273 % (Santali), and the median language pays 75 %. Across all 6 models on this page, the only exceptions are Chinese (Simplified): DeepSeek −6 %, Qwen −9 %; Yue Chinese: Qwen −3 %.
Why does Chinese pay less on Chinese models?
Chinese (Simplified) pays 25 % on ChatGPT / Copilot, but −9 % on Qwen and −6 % on DeepSeek. Alibaba (Qwen) and DeepSeek built their tokenizers from much more Chinese text, so common Chinese words get tokens of their own. A tokenizer is cheapest for the languages it was built for.
What is the token tax for the biggest languages?
The most spoken languages, and Swedish, on all 6 models:
| ChatGPT / Copilot | Gemini | Claude | Meta AI / Llama | DeepSeek | Qwen |
|---|
| Chinese (Simplified) | 25 % | 8 % | 14 % | 6 % | −6 % | −9 % |
| Spanish | 32 % | 29 % | 45 % | 30 % | 50 % | 31 % |
| Hindi | 57 % | 30 % | 116 % | 65 % | 180 % | 98 % |
| Standard Arabic | 38 % | 45 % | 81 % | 59 % | 62 % | 30 % |
| Portuguese | 23 % | 25 % | 37 % | 23 % | 42 % | 23 % |
| Russian | 42 % | 35 % | 32 % | 23 % | 58 % | 35 % |
| Japanese | 66 % | 21 % | 26 % | 31 % | 42 % | 16 % |
| French | 37 % | 40 % | 48 % | 37 % | 55 % | 38 % |
| German | 31 % | 32 % | 72 % | 30 % | 51 % | 30 % |
| Swedish | 35 % | 36 % | 32 % | 32 % | 55 % | 37 % |
Which AI model has the lowest token tax?
Measured by the median language: ChatGPT / Copilot 75 %, Gemini 81 %, Claude 84 %, Meta AI / Llama 86 %, Qwen 90 % and DeepSeek 107 %.
Which languages pay the least?
English, which the token tax is measured against (0 %). Only Chinese on Chinese models pays even less, see above. After English, the five cheapest on ChatGPT / Copilot are Portuguese 23 %, Dutch 25 %, Chinese (Simplified) 25 %, Indonesian 25 % and Standard Malay 30 %.
Which languages pay the most?
The native-speaker figures are from Wikipedia (census and Ethnologue figures). The 10 most expensive of 203 languages on ChatGPT / Copilot, with their token tax and number of native speakers:
- Santali 1 273 %: about 7,600,000 native speakers, mainly in India, Bangladesh, Nepal; Ol Chiki script.
- Tamasheq 921 %: about 900,000 native speakers, mainly in Mali, Burkina Faso; Tifinagh script. Uncertain figure.
- Central Atlas Tamazight 915 %: about 2,700,000 native speakers, mainly in Morocco; Tifinagh script. Uncertain figure.
- Dzongkha 803 %: about 171,000 native speakers, mainly in Bhutan; Tibetan script. Uncertain figure.
- Lao 723 %: about 3,700,000 native speakers, mainly in Laos; Lao script.
- Standard Tibetan 713 %: about 1,200,000 native speakers, mainly in China (Tibet), India, Nepal; Tibetan script. Uncertain figure.
- Shan 693 %: about 4,700,000 native speakers, mainly in Myanmar, Thailand; Myanmar (Shan) script. Uncertain figure.
- Tigrinya 500 %: about 9,900,000 native speakers, mainly in Eritrea, Ethiopia; Ge'ez script.
- Amharic 482 %: about 31,800,000 native speakers, mainly in Ethiopia; Ge'ez script.
- Odia 399 %: about 34,000,000 native speakers, mainly in India (Odisha); Odia script.
Do languages with other writing systems pay more?
Not on average: on ChatGPT / Copilot, the median language in the Latin alphabet pays 72 % and the median language in another script 79 %. But the most expensive languages all have scripts of their own, such as Santali (Ol Chiki), Tamasheq (Tifinagh), Dzongkha (Tibetan), Lao and Shan (Myanmar).
How can I measure the token tax of my own text?
Count the tokens of your text and of its English translation with the tokenizer of the model you use, and divide the first by the second. OpenAI's tokenizer is public; Anthropic and Google count tokens through their APIs. One sentence gives a rough number; the more text, the more reliable.
By Sara Moricz, research data scientist. Updated 2026-10-26.