Token

A token is the smallest unit a language model uses to process text — it doesn't equal a character, a word, or a sentence. English tokenization roughly follows this logic: common words are one token (cat = 1 token), while uncommon words get split (tokenization = 4 tokens). Chinese characters are typically 1-2 tokens each, because tokenizers were historically optimized for English. The most practical reason to understand tokens is cost: Anthropic API billing is by token, not by character count, not by message count. The tokens in your prompt plus the tokens in Claude's response equals the cost of that API call. A 1,000-character Chinese article is roughly 1,500-2,000 tokens; the same information in English is about 1,300 tokens — Chinese costs 20-50% more than English.

At its core, a language model operates on vectors (embeddings), and the input unit is a token, not a character. If characters were used, the vocabulary would explode in size, making it impossible for the model to learn effectively. Tokenization is an engineering trade-off: using common subword units allows processing of all languages while keeping the vocabulary manageable (typically 50,000-100,000 tokens). English has the most training data so tokenizers are most efficient for English; CJK languages have lower token efficiency, directly affecting API costs.

Token counts affect three key decisions: Cost estimation: Before building an API application, estimate tokens in your prompt and Claude's expected response, multiply by the API unit price. Don't use character counts — for Chinese text, character count times 1.5-2 gives a closer approximation. Context Window allocation: Every token in your System Prompt is billed again on every API call. 2,000-token System Prompt times 10,000 calls per day = 20 million input tokens per day just from the System Prompt. Streamlining it is the most direct cost optimization lever. Language selection: If your application is bilingual, the same content in English uses 20-50% fewer tokens than in Chinese. Many developers write System Prompts in English even for Chinese-facing products — a reasonable cost efficiency decision.

Three immediately actionable steps: Use Anthropic's token counting tool: In the Console Workbench, entering your prompt directly shows the token count. Calculate your cost structure before deploying — don't discover unexpected costs after launch. Compress your System Prompt: Remove redundant explanations and examples. Replace descriptive text with directive text. Reply in Traditional Chinese uses more than twice fewer tokens than a verbose equivalent instruction. Track token usage: Log the input/output token counts from the usage field in every API response. When a conversation's input tokens exceed a preset threshold, trigger a prompt for the user to start a new conversation, or automatically compress conversation history.