AI Tokens Explained: Understanding How LLM Billing Works

What Are Tokens?

Tokens are the basic units that large language models use to process text. Think of them as "text fragments" — not characters, not words, but something in between.

For English, one token is roughly 4 characters or 0.75 words. For example, "Hello world" is 2 tokens. For Chinese, one character is typically 1-2 tokens. For example, "你好世界" is approximately 4-6 tokens.

Why Bill by Tokens?

The computational cost of LLMs is directly proportional to the number of tokens processed. The model performs mathematical operations on each token — more tokens means more computational resources needed. Token-based billing is therefore the fairest pricing model.

Input Tokens vs Output Tokens

AI model billing typically distinguishes between input and output:

• Input Tokens (Prompt): Text you send to the AI, including system prompts and conversation history
• Output Tokens (Completion): Text generated by the AI in response

Output tokens typically cost 2-4x more than input tokens, as generating text requires more computation than understanding it.

Estimating Token Usage

Some rules of thumb:

• 1,000 English words ≈ 1,300 tokens
• 1,000 Chinese characters ≈ 1,500-2,000 tokens
• One A4 page of text ≈ 500-800 tokens
• A typical conversation turn (one Q&A) ≈ 500-2,000 tokens

Factors Affecting Token Consumption

Context Length

In every conversation, the model processes the complete conversation history. This means each message consumes more tokens as the conversation progresses.

System Prompts

System prompts are sent with every request and consume tokens. Keeping your system prompts concise can save costs.

Model Selection

Token prices vary significantly between models. Smaller models (like GPT-4o Mini) might cost 1/20th of larger models. For simple tasks, using smaller models can dramatically reduce costs.

GPTGet Billing

GPTGet offers flexible billing options:

• Subscription: Fixed monthly fee with included token allowance
• Pay-as-you-go: Purchase token packages, buy more when depleted

On GPTGet's usage statistics page, you can monitor token consumption in real-time and understand the cost of each conversation.

Token-Saving Tips

1. Be concise: Avoid unnecessarily lengthy descriptions
2. Start new conversations: When topics change, start fresh to avoid carrying irrelevant history
3. Choose the right model: Use smaller models for simple tasks, larger ones for complex tasks
4. Use workflows: Batch processing is more efficient than one-by-one processing