A journalist conducted a field test on an AI product from a telecom operator, spending CNY 9.9 to purchase 5 million tokens. After sending a simple "hello," the backend replied, "Hello! How can I help you?" This single, simple interaction consumed 49,000 tokens, as shown by the backend. Subsequently, the journalist issued a command to create a July schedule table, which directly consumed 500,000 tokens. In less than 30 minutes, the 5 million tokens were quickly depleted.
01. What Is a Token, and Why Do Pleasantries Cost You Money?
A token is the smallest data unit that AI (Artificial Intelligence) can read and generate, which can be understood as a word piece. While humans speak word by word, AI first segments the input text into several semantic fragments, and these fragments are tokens. Roughly converted, one Chinese character equals about 1.5 tokens. AI charges based on tokens; what you pay for is not the word count, but computing power.
Zhu Bo, Deputy Director of the Data Resource Center at Wuhan Data Group, explained that users are not merely purchasing the output word count, but the computing power resources consumed behind the scenes. During the input phase, your question is segmented into a token sequence and fed into the large language model; during the output phase, the model generates the response token by token, and each generated token requires complex calculations based on all preceding tokens. The more tokens there are, the longer the GPU computing time, and the higher the cost.
This explains why a single "hello" can burn through tens of thousands of tokens; it is not just charging for the two characters of "hello." In every round of conversation, the system must include the complete conversation history, context information, and tool call results to enable the model to make a coherent response. The pleasantries you mentioned earlier will be repeatedly calculated in every subsequent round of conversation.
02. How Much Can You Actually Save by Dropping Pleasantries?
Someone conducted an experiment by adding 8 lines of rules in Claude Code, with the core principle being simply "no nonsense." No flattering opening remarks, no polite closing blessings, no restating the question, just give the answer directly. As a result, the total output tokens decreased by 63% with zero information loss. However, it should be noted that these rules themselves must be loaded as input in every round of conversation, yielding a net benefit only when used at high frequency. Another open-source plugin called Caveman claims to save an average of 65% of output tokens, but this applies only to the output part, leaving the input tokens unchanged.
JetBrains' field tests based on 86 real-world software engineering tasks showed that the actual output tokens decreased by about 8.5%, far below the claimed 65%. To simply understand the above content: if you ask AI to write a paragraph about spring, typing "Hello, could you please help me write a paragraph about spring, thank you" versus directly typing "Write a paragraph about spring," the latter might only require one-third of the tokens needed for the former.
03. These Three Good Habits Can Help You Save Money
The first good habit is to break down complex tasks into smaller questions rather than dumping everything at once. Many people are accustomed to writing an entire paragraph of background, requirements, and constraints into a single message. However, tokens in the context are billed cumulatively; all the content you wrote previously must be re-read by the AI every time it generates a single character. Breaking a complex task into several rounds of short conversations is more economical than submitting a lengthy essay all at once.
The second good habit is to use cheaper models for daily chitchat and reserve expensive ones for complex problems. The unit price of tokens for different models can vary by dozens of times. Use the standard mode for checking the weather or looking up words, and switch to high-performance models for in-depth analysis and professional writing. Additionally, do not restate the question. AI does not need you to repeat the question before making a request. Stating directly what you want saves much more than saying, "I would like to ask a question, it is about... could you please help me..."
04. Final Words
Tokens are not just the pricing unit when you chat; they are becoming the universal currency for AI services. In May 2026, the three major telecom operators simultaneously launched token packages, with a minimum of CNY 9.9 for a quota of 10 million tokens. Wuhan Data Group is integrating government data such as business registration, social security, real estate, and medical insurance to build an AI government large model. In the future, when citizens apply for business licenses or check their housing provident fund, they only need to ask questions in natural language, and the AI will provide answers on the spot. The settlement between the government and the operators will be calculated based on the actual amount of tokens consumed.
Tokens are being measured and traded just like water and electricity. Every time you ask AI a question, behind it is a measurable and priceable consumption of computing power. Next time you open an AI dialog box, try deleting words like "hello" and "may I ask," and state directly what you want. This habit itself is saving you money.
#AI