Loading...
How AI models split text into billable chunks. Tokens are the meter on every LLM bill.
Quick Definition
Tokenization is the process by which AI language models break text into tokens, small chunks of characters or words, which are the units in which usage is measured and billed. Because most LLM APIs charge per token, tokenization is the meter that drives AI API costs.
Tokenization is the process by which language models break text into tokens, small chunks of characters that are roughly three quarters of an English word on average. Models read and generate text token by token, and because LLM pricing charges per token, tokenization is literally the meter that measures AI API spend.
The cost implications are concrete. Everything sent to a model counts: the user's question, the system prompt, conversation history, and any attached documents all become input tokens on every call. A chat application that resends the full conversation with each message pays for the same tokens again and again as the thread grows.
Example. A 1,000-word document is roughly 1,300 tokens. An app that attaches it to all 50 calls in a session pays for about 65,000 input tokens of the same content. Caching or summarizing that context cuts the repeated cost by more than 90 percent.
For FinOps teams, tokens are a new metered resource to govern like compute or storage: measure consumption per feature, set token budgets, alert on anomalies such as runaway loops, and report cost per request. The token budgeting guide is a practical starting point, and FinOps for AI places token control inside the wider AI cost discipline.
Roughly 1.3 tokens per English word on average, so 1,000 words is about 1,300 tokens. Code and non-English text often tokenize less efficiently.
Yes. Input tokens (prompts, history, documents) and output tokens (responses) are both billed, with output usually priced higher.
Long contexts resent on every call, runaway agent or retry loops, and verbose outputs. Caching, trimming, and per-feature budgets catch all three.