When an API bills by token, the language of a prompt can affect its cost. Equivalent content in English, Chinese, Korean, or another language may produce different token counts because providers use different tokenizers.
The comparison below, attributed in the original post to researcher Aran Komatsuzaki, illustrates that difference for selected models and texts. It is a snapshot, not a universal “language tax” for every model or conversation.
The original chart showed higher Chinese token counts for a Claude model and more favorable Chinese tokenization for some models developed in China. This does not establish that a model lacks non-English understanding; token efficiency and language quality are different questions.
Measure the text you actually send using the provider's tokenizer or usage report. Include outputs and repeated context when estimating costs. Writing a prompt in English may reduce its token count in some cases, but translating it can also change its meaning or increase the response length.
Adapted from the original Chinese article, published on May 5, 2026.
Comments NOTHING