Input vs Output Tokens: Why the Split Matters
AI pricing guide · updated 2026-08-07
Every AI API quotes two prices, and most comparisons quietly average them into one. That average hides the decision that actually matters: whether your workload is dominated by what you send or by what the model writes back. Get the split wrong and the "cheaper" model costs you more.
The asymmetry, in numbers
Across the 271 text models we track, the median charges 2× more per output token than per input token. 117 of them charge at least 3× more. The extreme is 8.3×, and only a handful price the two sides equally.
So a single blended figure is a convenience, not a measurement. It assumes a token mix, and if your mix differs from the assumption the ranking it produces can be wrong.
Why generation costs more
Input is consumed in one forward pass, with all tokens attended to in parallel. Output is produced sequentially: each new token requires another pass over the entire context, including everything generated so far.
That makes output fundamentally more expensive to serve, and it also explains why reasoning models cost so much more in practice. A model that "thinks" before answering is generating a long internal chain of tokens, and you are billed for those at the output rate even when they never reach your screen.
Three workload shapes, three different winners
Input-heavy work (retrieval-augmented generation, document analysis, classification of long text) sends a lot and returns a little, often 80/20 or steeper. Here the input rate dominates and a model with cheap input wins even if its output rate is poor.
Output-heavy work (long-form drafting, code generation, agent loops) inverts that, sometimes to 20/80. Here the output rate is nearly the whole bill, and the cheapest-on-average model can be a poor choice.
Balanced work (most chat and assistant use) sits near the middle, though bear in mind a growing transcript pushes chat steadily toward input-heavy over a long session.
This is why our rankings are published on several weightings rather than one: the cheapest model for retrieval is frequently not the cheapest for generation.
How to find your own ratio
Log the input and output token counts your provider returns on each response and average them over a representative day. Most APIs report both in the usage field of every response, so this needs no estimation.
Then price candidate models against those two numbers separately rather than against a blended figure. The arithmetic is trivial and the answer is frequently different from the one a single-number comparison gives.
If you have not shipped yet and have no logs, estimate deliberately conservatively on output: it is the side that surprises people, and it is the expensive side.
Frequently asked questions
What is the difference between input and output tokens?
Input tokens are everything you send: system prompt, retrieved context, conversation history and the user message. Output tokens are everything the model generates in response. They are billed at separate rates, and output is almost always the more expensive of the two.
Why are output tokens more expensive?
Because generating them costs more compute. Input is processed in a single parallel pass, whereas each output token requires its own pass over the whole context. The median model we track charges 2× more for output, and some charge over 8× more.
Does a blended price tell me what I will pay?
Only if your token mix matches the assumption behind it. A blended figure weights the two rates by a fixed ratio (commonly 70% input, 30% output) which is useful for ranking but not for budgeting. Price your own measured input and output volumes separately for an accurate answer.
Do reasoning tokens count as output?
Yes. Models that reason before answering generate intermediate tokens that are billed at the output rate even when they are hidden from the final response. That is why reasoning models can cost several times more than their headline rate suggests for a given task.
See it in the data
Related guides
Terms used in this guide
Published by Tokenando. Last updated 2026-08-07. Figures in this guide are computed from our own pricing index and dated where they can move; see the methodology and corrections policy.