APIs bill by the token and models think in tokens. A token is a word chunk: common English words are one token, rarer ones split into pieces, and many Chinese characters take one to two tokens each. Paste text to estimate the token count under different model tokenizers.
The estimation rules
GPT-4o and the newer o-series tokenizer (o200k): English ≈ 0.25 tokens per character, Chinese ≈ 0.7 tokens per character. The older GPT-3.5/GPT-4 tokenizer (cl100k): English ≈ 0.25 per character, Chinese ≈ 1.5–2 per character. Claude lands between. Expect ±10% error — precise counts need each vendor's tokenizer, but estimates are fine for budgeting.
Why Chinese costs more
English words are pre-chunked into common tokens; Chinese characters get split into sub-word pieces. The same 1,000-word essay can cost two to three times more tokens in Chinese than in English. If a task works in English, writing it in English literally costs less per run.
Context windows in plain English
The context window is the total budget for your prompt plus the entire conversation history plus the reply. GPT-4o offers 128K tokens; Claude's long-context models offer 200K — roughly 100,000+ words of English. Exceed it and the model starts forgetting the earliest parts of the conversation.