Almost every provider charges output tokens three to five times more than input tokens. It is the most important asymmetry on the bill and the one most often ignored: people optimize the prompt, which is the cheap part, and leave the answer untouched, which is the expensive one.
If it is not yet clear what is being counted, start with what a token is and how it is billed.
At a glance
| What happens | Output costs 3 to 5 times more than input |
| Where it shows | Long answers, reasoning, free-form output |
| The lever | Shorten and structure the answer |
| When it pays off | Whenever output weighs on your spend |
Why output costs more
Input is processed in one go: the model reads the whole prompt in parallel. Output is generated token by token, and each one means running through the entire model again. It is sequential work and it does not parallelize as well, so the provider charges more for it. The difference is not a commercial whim, it reflects the real cost of generating.
When output dominates the bill
There are three typical cases. Long answers by default, when nobody has set a limit and the model rambles. Reasoning models, which produce intermediate tokens you pay for even though you never see them. And content generation, where the job is precisely to write a lot.
If instead your load is classifying, extracting fields or answering over long documents, the weight is on the input and the lever is a different one: the cost of long context and prompt caching.
How to trim the expensive half
Set an explicit ceiling. An output token limit per task type stops one request from running away. Without it, the worst case cost of a request has no cap.
Ask for structure. An object with four fields takes a fraction of what a paragraph explaining the same thing takes, and it is easier to use downstream.
Drop the pleasantries. Preambles along the lines of “sure, here is the analysis you asked for” are billed tokens that add nothing.
Separate thinking from writing. If you need the model to reason, let it reason where it is cheap and write where it matters, instead of paying for reasoning at the large model’s price.
Put a number on it
Take a week of real traffic and split the spend between input and output. With those two numbers and the token cost calculator you will know whether your problem is the prompt or the answer before touching anything. It is the first cut of the cost per task.
Frequently asked questions
How much more expensive is output? It depends on the provider and the model, but the usual range is three to five times the input price.
Do reasoning tokens get billed? Yes, they are billed as output even though they never appear in the final answer. On reasoning models they can be most of the cost.
Does capping output hurt quality? Only if the cap falls below the useful answer. Measured per task there is almost always room: long answers are rarely the best ones.
At e-ficient we measure the cost per task of companies that already run AI in production and tell you how much you would save with each lever at your real volume. The first audit is free if you start a plan with a three-month commitment and returns a diagnosis within 72 hours.