AI Bloat: the waste you pay for on every call

What AI Bloat is

AI Bloat is all the content we send to an AI system that consumes budget without changing the result. The broad definition covers a lot —retrieved context you did not need, history that no longer matters, tool definitions never used— but a definition that covers everything cannot be measured.

So we work with a narrow one, which is the one that can actually be checked: the text you can remove from a prompt without altering what it asks for. Everything below refers to that, and every figure we give is measured against it.

The name comes from software bloat, the same phenomenon in programs that grow without anyone deciding they should. A prompt grows the same way: someone adds a sentence because the model failed once, the sentence works, and it stays there forever.

Why it matters: it multiplies

Waste in a prompt is not paid for once. The fixed part —the system prompt, the instructions that always travel— is sent on each and every call your product makes.

A thousand spare tokens in a product making ten thousand calls a day is ten million tokens a day that have nothing to do with what any user does. Trimming 30 % off that fixed part does not save 30 % of one call: it saves three million tokens a day, every day, without touching the quality of a single answer.

And if you do not pay per API call, it still costs you. Every token takes up context window and brings forward the moment the conversation gets cut or the model starts forgetting the beginning. It is the same waste billed in a different currency.

How it is measured

The basic measurement is the bloat rate: the share of tokens that can be removed without changing what the prompt asks for.

It is worked out by counting the tokens of the original text, applying a known rule set, and counting the tokens of the result. The difference, divided by the original, is the rate. That is all.

What makes that figure worth anything is that it is deterministic: the same prompt gives the same number today and in a year, and anyone can repeat the operation and get the same result. A tool that rewrites by calling a model cannot give you that, because its answer changes between runs — and if the result changes, there is no way to know how much you saved.

• ·It is measured on the text without code, URLs or variables: those are neither waste nor instruction, so they stay out of the calculation instead of contaminating it.
• ·The exact count comes from the model's own tokenizer. A character estimate is good for an order of magnitude, not for a decision.
• ·Rates compare within one language. Spanish fragments worse and carries more roundabout phrasing than English, so comparing across languages says nothing.

What the waste is made of

Not all spare text is spare in the same way, and the distinction matters because the risk of removing it is not the same. These six categories run from least to most invasive.

• ·Pleasantries: greetings, «please», thanks and sign-offs. The model does not obey better for politeness. It is the largest category in prompts written by people.
• ·Filler: verbal tics and reinforcement such as «it is very important that» or «always remember». If the instruction is already an imperative, reinforcement adds no obedience.
• ·Verbosity: roundabout phrasing that can be said shorter. «In order to» for «to», «in the event that» for «if». It changes the length, not the meaning.
• ·Redundancy: the same rule stated twice in different words. It is the typical residue of months of patching, and usually what takes up most room in a veteran system prompt.
• ·Structure: inflated role definitions. «You are a financial analyst» does the same job as four lines describing their experience.
• ·Formatting: decorative separators, deep indentation and spare whitespace. It looks like organisation and is billed as text.

What this measurement still cannot see

This is worth saying before anyone else does, because it decides whether a measurement can be used to make decisions or only to be shown.

The rate measures text a rule set can recognise and remove. It does not detect that two paragraphs say the same thing in different words: «answer in Spanish» and «remember you have to reply in Spanish» are, to a literal comparison, two different sentences. A system prompt heavily inflated by reworded repetition can score a low rate and still be full of waste.

It does not measure quality either. A rate of zero does not mean the prompt is good: it means it has no recognisable fat. You can write a terrible prompt that is perfectly compact.

And someone decides the rules. That is the most reasonable criticism you can make of any measurement of this kind, and the only valid answer is to publish the rules so they can be argued with one by one.

Why it is not prompt engineering or «token optimization»

Prompt engineering is a craft: it aims to make the model do what you want. AI Bloat is a measurement: it says what asking costs you. They are independent axes, which is why a prompt can be excellently designed and enormous at the same time.

«Token optimization» names the solution, which is what a vendor does when describing its product. AI Bloat names the problem, which is what the person suffering it does. The difference is not stylistic: nobody ever bought «refactoring», they buy because they have technical debt.