How to calculate GPT words?

When you deposit money into your balance, we calculate how many words it should be. The GPT charges by tokens, which we convert to words for clearer understanding.

What are GPT Tokens?

GPT tokens are like building blocks that a computer uses to understand and create text. Imagine you have a box of Lego blocks, where each block represents a word or a part of a word. When you build something with Legos, each block has its place to make the final model look right. Similarly, GPT looks at these blocks (tokens) of language to understand what you mean or to help it write sentences that make sense. So, when you talk to GPT, you’re giving it these blocks, and it uses them to build back sentences or answers for you.

How does GPT calculate spent tokens?

GPT calculates the number of tokens spent based on the length of the text it processes, including both the input (what you ask or provide to it) and the output (the response it generates). Here’s a simple breakdown of how it works:

  1.  **Tokenization**: GPT first breaks down the text into tokens. Tokens can be words, parts of words, or even punctuation marks. The exact nature of a token depends on the model’s training and its tokenization algorithm.
  2. **Counting Tokens**: Once the text is tokenized, GPT counts the number of tokens in both the input and output. The total number of tokens used is the sum of input and output tokens.
  3. **Spending Tokens**: Each token that GPT processes (reads or writes) counts towards the total tokens spent. If you have a limit on the number of tokens you can use, this total will be deducted from your available tokens.

For example, if you ask GPT a question and the question (input) breaks down into 10 tokens, and GPT’s answer (output) breaks down into 20 tokens, you’ve spent a total of 30 tokens on this interaction.

The model’s efficiency and your usage costs depend on how concisely you phrase your inputs and how you configure the model to respond (e.g., setting limits on response length).

 

How many tokens does one word represent?

Think of GPT as a machine that breaks down sentences into smaller parts, called tokens. Different languages are like different types of bread. Just like slicing bread, some breads are denser, so you get fewer slices from a loaf, while others are lighter, giving you more slices.

Here’s a simple way to see it:

  • **English** is like white bread; one word usually splits into about 1.3 tokens.
  •  **French, German, and Spanish** are like whole grain bread; one word gets about 2 to 2.1 tokens.
  •  **Chinese** is like sourdough; dense, so one word equals about 2.5 tokens.
  • **Russian and Vietnamese** are like pumpernickel; very dense, stretching one word to 3.3 tokens.
  •  **Arabic** is like rye bread; one word can be about 4 tokens.
  • **Hindi** is the densest, like a fruit loaf; one word goes up to about 6.4 tokens.
  • ..etc

These comparisons help understand how GPT “slices” different languages into tokens. Some languages “cost” more tokens because of how they’re split up, just like some breads give fewer slices from a loaf.

How do we calculate the words count for your balance?

We provide an estimated word count because the actual calculation depends on the GPT model and language you use. The calculation varies by GPT model and language. For example, in our calculation, one English word equals 1.3 GPT tokens. Therefore, 1000 English words equal 1300 GPT tokens. The cost for 1000 tokens is $0.005, making the price $0.0065 for 1000 English words in GPT-3-Turbo mode. For GPT-4-Turbo, the price will be $0.06*1.3 = $0.078 for 1000 English words.