Tokens are a basic building block used by generative artificial intelligence (AI) models to process and understand information. These models break down text, images, and other data into smaller units called tokens, which can represent a full word, part of a word, punctuation, or a few characters. The way text is divided into tokens depends on the specific tokenizer used, and different models may split the same sentence differently. In English, a token often corresponds to about four characters or three-quarters of a word, though this can vary by language and the tokenizer's design. Tokens serve two main purposes in AI systems: they are a unit for processing information and a way to measure the model's capacity and usage. The context window, or the amount of information a model can handle at one time, is often measured in tokens. For example, a model with a context window of 200,000 tokens can process that many tokens simultaneously. However, this doesn’t mean the model remembers all that information after the task is complete. When interacting with an AI, the input tokens include the user's question, previous messages, instructions, and any document content. Output tokens are the model's generated response. Some models also use cached tokens, which are input tokens that have been processed before and reused to save resources, and reasoning tokens, which are used during intermediate steps of processing but may not appear in the final output. The cost of using AI models is often based on the number of tokens processed. Input tokens are usually less expensive than output tokens, as generating each output token requires more computational power. Pricing can vary widely between different models and service providers, with some models charging several dollars per million tokens. Multimodal models, which handle text, images, audio, and video, also use tokens to measure consumption, but the conversion of these media into tokens is not standardized and can vary based on the model, resolution, and processing method. Developers often provide tools to help users estimate the number of tokens used in their interactions.