Token
The basic unit an LLM reads and generates — roughly a fraction of a word.
LLMs don't process raw characters or words. They process tokens — subword fragments produced by a tokenizer. "Hello world" is roughly 2 tokens; "unbelievable" is 3-4 tokens. Across English text, 1 token averages ~0.75 words.
Tokens are the unit of pricing for LLM APIs (input tokens cheaper than output tokens, generally), the unit of context window measurement (200K context = 200K tokens), and the unit of speed measurement (tokens per second).
For operators: when you see pricing like "$3 per million input tokens," do a back-of-envelope check on your actual usage. A single agent run that processes a 50K-token codebase and outputs 5K tokens of analysis costs under $0.50 at frontier-model rates. That math is what makes most agent products economically viable.
Get the weekly digest
New tools, reviews, and prompts every Friday.