Every language but English pays more for AI

• +56%more tokens than English on average across the other 24 languages, on OpenAI's current tokenizer.
• +21%more tokens on DeepSeek's tokenizer than on OpenAI's across all 25 languages, against only 2% on English, the language tokenizers are usually compared in.
• 54%of the extra cost disappeared on average when OpenAI moved from the GPT-4 tokenizer to GPT-4o's. The rest is still there.
• 2.32×Greek is the most expensive of the 25 languages measured: the same text costs 2.32 times as much as in English.

Claude, Gemini and Grok are not measured yet; their model pages use estimates until they are. Anthropic documents that Claude 4.7 and later use about 30% more tokens than earlier Claude models for the same text.

1. 1

   Write system prompts and fixed instructions in English, and keep only the user-facing text in your language. Models follow English instructions just as well, and those instructions are sent with every request.

2. 2

   Compare models on text in your own language, not on English. A model that is cheaper per token can cost more per sentence.

3. 3

   Check the tokenizer, not only the price. OpenAI's newer tokenizer narrowed the gap, while the one in Claude 4.7 uses more tokens than its predecessor. Newer does not automatically mean cheaper.