Managing LLM cost is a basic LLMOps skill. Let us cover: 1) How cost is counted: you pay for TOKENS (pieces of text, roughly parts of words). Two kinds: — Input tokens: your whole prompt — instruction, system message, context, data, dialogue history. — Output tokens: the models answer (usually costlier per token). Request cost = (input tokens + output tokens) at the models rate. Total for a period = request cost × number of requests. At volume (thousands-millions of requests) even a small per-request cost adds up to large sums. 2) What inflates cost (common causes): — Too powerful a model for a simple task: using a top expensive model where a compact cheap one copes — overpayment. — Long prompts: redundant instructions, excess context. In RAG — too many/long fragments in the context. — Long dialogue history: passing the whole history each time — growing cost. — Repeated identical requests: paying for what was already computed (solved by a cache). — No limits: unrestricted requests, abuse. 3) How to reduce cost: — Model choice for the task: use the cheapest model that copes. For the simple — a compact one; a powerful one only where truly needed. Often this is the main saving. — Shorter prompts: remove the excess from instructions and context. In RAG — feed only relevant fragments, not the whole document. — Manage history: do not pass the whole dialogue history if not needed (summarize the old, trim). — Caching: save answers to repeated requests (next lesson). — Limit output length: if you do not need long answers, limit it. — Per-user/request limits: protection from abuse and unexpected spikes. 4) Tracking: it is important to SEE the cost — how much you spend, on what, which requests are costlier. Without tracking you learn of a problem from a big bill. Cost monitoring (part of general monitoring) lets you notice and optimize. 5) Forecasting: estimate the cost at scale IN ADVANCE. "Cost per request × expected volume" — does it make your products economics viable? Better to understand before launch than after a big bill. Practical rule: understand how cost is counted (input/output tokens × volume), choose the cheapest suitable model, shorten prompts and context, cache repeats, set limits, track spending and forecast at scale. Cost control is not a one-off action but a constant practice. Uncontrolled cost is one of the main reasons LLM projects become unprofitable in production. Start thinking about cost from the start, not when a big bill comes.