Does saying please to an LLM cost money?

A handful of tokens per call, which is real but trivial. The expensive parts of a prompt are redundant instructions, stale defensive boilerplate and duplicated few-shot examples, not courtesy.

Why is prompt bloat worse in agents?

A single user action can trigger many model calls, and the system prompt is resent on each one. The same fixed overhead is billed once per call rather than once per task.

How do you safely trim a prompt?

Run an ablation against an eval suite: remove one block at a time, re-run the evals, and keep the removal only if quality holds. Trimming without evals risks a regression that costs more than the tokens saved.

Quick answer: Prompts grow the way configuration files grow. Someone adds an instruction to fix a failure mode. Someone else adds an example. A third person adds a paragraph explaining why the second instruction...

How prompting style shows up on the bill

Updated September 1, 2026 · first published September 1, 2026

Prompts grow the way configuration files grow. Someone adds an instruction to fix a failure mode. Someone else adds an example. A third person adds a paragraph explaining why the second instruction matters. Nothing is ever removed, because removing it might reintroduce a bug nobody can reproduce. Six months later the system prompt is three thousand tokens and no one can say which two hundred of them are load-bearing.

Why it compounds

Unlike output, which varies with the task, prompt overhead is paid on every single request, identically, forever. A thousand wasted tokens on a feature serving ten million calls a month is ten billion tokens a month of pure formatting.

In multi-turn conversations it is worse: the system prompt is resent on every turn, so its cost scales with turn count, not request count. And in agent loops, where a single user action can trigger a dozen model calls, the same bloated preamble is billed a dozen times for one piece of work.

What is actually wasteful

Not politeness, which is a handful of tokens and occasionally improves compliance. The real weight sits in: redundant instructions that restate the same rule three ways, defensive boilerplate added for a model generation that no longer needs it, examples that duplicate rather than cover distinct cases, and explanations aimed at human readers of the prompt file rather than at the model.

That third category is the largest in practice. Few-shot examples are expensive per token and teams rarely revisit whether five are doing work that two would do.

Treat the prompt as code with a budget

Measure the fixed prompt overhead as a share of average request size — if it is over half, it is a cost problem, not a style question. Put the stable part first so caching applies, which makes the remaining bloat far cheaper without deleting anything. Then run an ablation: remove one block, run your evals, keep the removal if quality holds. Prompt trimming without an eval suite is guesswork, and the failure mode — a quality regression discovered by customers — costs more than the tokens saved.

Want this applied to your own LLM spend? FinOps LLM runs a free audit of your AI costs and shows where the savings are. Book free audit →

Back to research