In this article＋
01 / ARTIFICIAL INTELLIGENCE
A token is neither a word nor a character
A common English word may fit one token while an unusual name breaks into several. Equal visible length does not imply equal model input.
You type 'make this shorter', but the request may carry thousands of tokens from the thread, files and tools behind it. One visible sentence, a whole bag backstage.
You type 'make this shorter', but the system may also carry the full thread, an uploaded file, search results and tool output. The visible prompt is tiny; the working context is not.
02 / ARTIFICIAL INTELLIGENCE
The split is part of model design
The vocabulary comes from a design and training process. Frequent fragments are worth reserving as units; rare strings are decomposed.
Token count therefore reveals a hidden preference about which languages, scripts and spellings are familiar building blocks. It is not a neutral word counter attached after the model is complete.
The first saving is not wrestling with punctuation. Ask whether the old material is still needed.
For a Hong Kong team, the sensible saving is not shaving every word. It is sending only what the workflow needs, using caching and summarising long tasks. Cut noise, not necessary context.
03 / ARTIFICIAL INTELLIGENCE
The same meaning can have a different price
Multilingual research finds that equivalent information can require several times as many tokens in one language as another, depending on the tokenizer. Extra tokens consume context, add processing and may hit limits sooner.
A million-token window is not a million words and does not provide equal content capacity across languages.
Removing three pages of repeated background can beat trimming ten words from the prompt.
Input, cached input and output can have different rates, and tools or deeper reasoning can add usage. 'About this many words' is not a monthly budget.
04 / ARTIFICIAL INTELLIGENCE
Spaces, punctuation and formatting also count
JSON brackets and quotes, code indentation, table separators and chat-role markers all count. A short visible question may be accompanied by system policy, tool definitions, conversation history and retrieved documents.
The user's typed characters are only the visible corner of the token budget.
05 / ARTIFICIAL INTELLIGENCE
Output is often the expensive side
Services often price input, cached input and output separately. Output is decoded token by token, so latency and opportunity for error grow with length.
Be more detailed is not only a stylistic request. Useful savings usually come from removing repeated context and unnecessary output, not compressing a question into unreadable shorthand.
06 / ARTIFICIAL INTELLIGENCE
The only reliable answer is measurement
The familiar estimate of four characters per token is only a rough English guide. Run the exact prompt through the tokenizer associated with the chosen model, including system messages, history, tool schemas and expected output.
The hidden truth is that you pay not only for the question, but for the entire conversational stage around it.
Token efficiency is not writing quality
A shorter token sequence is not automatically a clearer prompt, and the language using fewer tokens is not automatically understood better. Compression, training coverage and task performance interact.
Optimising only the count may remove examples or structure that prevent an expensive mistake. Measure cost per successful task rather than cost per request: a prompt that uses twenty percent more context but avoids a second run may be the cheaper system.
07 / ARTIFICIAL INTELLIGENCE
The cost often hides in what you do not count as content: history, tool output, formatting and repeated instructions.
You type 'make this shorter', but the system may also carry the full thread, an uploaded file, search results and tool output. The visible prompt is tiny; the working context is not.
For a Hong Kong team, the sensible saving is not shaving every word. It is sending only what the workflow needs, using caching and summarising long tasks. Cut noise, not necessary context.
08 / ARTIFICIAL INTELLIGENCE
Token counts are not a fixed price
Input, cached input and output can have different rates, and tools or deeper reasoning can add usage. 'About this many words' is not a monthly budget.
Use real samples instead: one long file, one ordinary question and one tool-using workflow. Measure them separately before scaling.
09 / ARTIFICIAL INTELLIGENCE
Questions people actually ask
Is Chinese always cheaper than English?
Not as a rule. Tokenisation depends on the tokenizer and is affected by punctuation, mixed language and formatting.
Does deleting history help?
If old material is no longer needed, it reduces context. Do not remove evidence that still supports a decision.
What is cached input?
Some platforms charge less when an identical prefix is reused. The exact rule depends on the product and model.
Sources
Sources
Sources support the mechanisms and limitations discussed here. Models, products and prices change; check the official page and date when a detail matters.