Switching languages in generative AI can lead to fewer tokens being used, according to research by companies like Anthropic, OpenAI, and Google. Tokens are the fundamental units that AI models use to process and generate text, and they vary depending on the model and language. Using fewer tokens can reduce costs, computational demand, and the environmental impact of AI operations. Each token requires the model to perform billions of calculations. These tokens are processed by the model and then returned in a similar format. The tokenizer, a tool that converts text into tokens, plays a crucial role in this process. The reverse operation of the tokenizer then turns these tokens back into the final output. However, not all tokens are the same; they are the basic unit for billing and resource allocation. The more tokens used in input and output, the higher the cost and the greater the environmental impact. The relationship between languages and tokens is significant because the way tokens are created depends on both the model family and the language being used. To illustrate this, the General Data Protection Regulation (GDPR), a lengthy regulation from the European Commission, was used as a reference. This regulation is translated into the 24 official languages of the European Union and serves as a good example of how language affects token usage. The text of the GDPR ranges in length from 320,000 to 415,000 characters, with an average of around 371,000 characters. This allows for the calculation of the average number of characters per token across different languages. It's important to note that the savings from using fewer tokens are relatively small compared to more impactful strategies. For example, avoiding tasks like image recognition or translation with general-purpose models like ChatGPT can lead to greater savings, as specialized AI systems are more efficient in these areas. Additionally, asking models to be more concise can significantly reduce token usage, leading to more substantial cost and resource savings.