Token lab: consumption optimization

In a real working session — a coding agent, a support bot, anything with tools and a history — the sentence you wrote is often well under one percent of what the model reads. The rest is the system prompt riding along on every single turn, the tool definitions, the file you attached six messages ago, and the conversation history compounding underneath all of it. The model has no memory, so every message re-sends the lot — and that packet is the bill.

So the first half is about seeing it. You'll take your own prompts apart and price them live, and leave able to read a prompt's composition before you send it: which slice is big, which slice is stable enough to cache, and which one is quietly doubling every turn. That is the difference between guessing at a cloud bill and predicting it.

The second half is the decision that saves the most money: which model does this job actually need? You'll leave with a working definition of "everyday" and "complex" — not opinions, but thresholds you measured yourself on testing and coding tasks — and a rule for when Haiku is genuinely enough, when Sonnet is the right default, and when Opus earns its five-times price tag. No vendor pitch, no prompt-engineering folklore.