Why it matters
Tokens are the meter running behind AI features: long documents in, long answers out, and every exchange in between all bill by the thousand tokens. Understanding it turns "why is this feature's AI bill spiky?" from mystery into arithmetic.
How it works
Text is split into common fragments, "translated" might be "transl"+"ated", and models process sequences of these. A page of text runs a few hundred tokens. Costs are quoted per million tokens, with output typically pricier than input.
What to do about it
For any AI feature at scale, estimate tokens per interaction × volume before building. Trimming what you send the model (summaries instead of full documents) is often the cheapest optimisation available.
