A journalist put an AI product from an operator to the test, buying 5 million Tokens for just 9.9 yuan. Upon sending a simple "Hello," the backend promptly responded with, "Hello! How can I assist you?"

This brief exchange consumed a staggering 49,000 Tokens, as revealed by the backend data. The journalist then asked the AI to create a schedule for July, which instantly used up another 500,000 Tokens. In under 30 minutes, all 5 million Tokens were exhausted.

01 What Exactly is a Token, and Why Do Pleasantries Cost Me?

A Token represents the smallest unit of data that AI can process and generate, akin to a linguistic building block. While humans articulate thoughts one word at a time, AI dissects input text into semantic fragments, which are the Tokens.

Roughly estimated, one Chinese character corresponds to approximately 1.5 Tokens.

AI's pricing model is based on Tokens, reflecting the computational power required, rather than the mere word count.

Zhu Bo, Deputy Director of the Data Resource Center at Wuhan Data Group, clarified that users are not just paying for the output words but for the computational resources utilized in the background.

During input, your query is segmented into Tokens and fed into the AI model. In the output phase, the model constructs responses Token by Token, necessitating complex calculations based on all preceding Tokens.

The greater the number of Tokens, the longer the GPU operation time, and consequently, the higher the cost.

This elucidates why a simple "Hello" can devour tens of thousands of Tokens—it's not merely about the two words themselves.

In each dialogue round, the system retains the entire conversation history, contextual cues, and tool call outcomes to ensure coherent responses from the model.

The pleasantries mentioned earlier undergo repeated calculations in subsequent dialogue rounds.

02 How Much Can I Save by Skipping Pleasantries?

An experiment was conducted using Claude Code, incorporating eight lines of rules aimed at eliminating unnecessary verbiage.

No flattering introductions, no polite sign-offs, no redundant questions—just direct answers. The outcome was a 63% reduction in total output Tokens without any loss of information.

However, it's crucial to note that these rules themselves must be loaded as input in each dialogue round, yielding a net benefit only with frequent usage.

Another open-source plugin, Caveman, claims to save an average of 65% on output Tokens, but this applies solely to the output portion—input Tokens remain unaffected.

JetBrains' tests, based on 86 real software engineering tasks, revealed an actual reduction in output Tokens of about 8.5%, significantly lower than the claimed 65%.

In essence, asking AI to write a paragraph about spring with a polite preamble, "Hello, can you please write a paragraph about spring for me? Thank you," versus a straightforward "Write a paragraph about spring," could result in the latter requiring only one-third of the Tokens.

03 Three Money-Saving Habits

The first habit is to break down complex tasks into multiple rounds of concise questions rather than overwhelming the AI with a single, lengthy message. Many are accustomed to cramming background information, requirements, and constraints into one paragraph.

However, Tokens in the context are billed cumulatively. AI must revisit all previously written content when generating each word.

Dividing complex tasks into several short dialogue rounds proves more cost-effective than submitting a verbose message all at once.

The second habit is to opt for cheaper models for casual conversations and switch to pricier ones for intricate problems. Token prices can vary significantly across models. Use standard mode for inquiries about the weather or word definitions, and reserve high-performance models for in-depth analysis or professional writing.

Also, refrain from repeating the question. AI doesn't require a preamble before processing your request. Stating your needs directly is more cost-effective than saying, "I have a question... about... Can you help me..."

04 Concluding Remarks

Tokens are not merely the pricing unit for your chats; they are evolving into the universal currency for AI services. By May 2026, the three major operators had all introduced Token packages, with a minimum of 10 million Tokens available for 9.9 yuan.

Wuhan Data Group is integrating government data, including business, social security, real estate, and medical insurance, to construct an AI government model. In the future, citizens will be able to apply for business licenses or inquire about housing provident funds through natural language queries, with AI providing instant answers.

Settlements between the government and operators will be based on the actual Tokens consumed. Tokens are being metered and traded like utilities.

Every query you pose to AI represents a measurable and billable consumption of computational power. Next time you initiate an AI dialogue, consider omitting words like "Hello" and "Excuse me" and state your request directly.

This habit alone can save you money.

#AI