Decompose the cost of a single API call into its components and explain which dominates.

Every LLM API call charges you for two kinds of tokens: the ones you send in (input) and the ones the model generates (output). The provider quotes a price per million tokens for each, and they are not the same price. Input tokens are cheap. The model reads your whole prompt in one parallel pass during the prefill phase, so the hardware is used efficiently. **Output tokens** are the expensive ones. The model produces them one at a time in the decode phase, which is slower and harder to parallelize, so providers charge roughly three to five times more per output token than per input token. There is a third category: cached input. If you reuse a long, stable prefix like a fixed system prompt, the provider can cache it and charge around ten percent of the normal input rate for those tokens on later calls. The surprising consequence is that output usually dominates the bill, even when you send far more input tokens than the model returns. A five times higher rate flips the ratio fast. So the highest-leverage cost lever is usually making the model say less, not feeding it less.

A single call has three priced components: new input tokens, cached input tokens, and output tokens. The total cost is: $$\text{cost} = n_{in} \cdot r_{in} + n_{cache} \cdot r_{cache} + n_{out} \cdot r_{out}$$ The rates differ sharply. Output runs about five times the input rate per token. Cached input runs about one tenth of the input rate. So a token is cheap to read, near-free to re-read, and several times pricier to generate. Output usually dominates on production workloads. The reason is the rate gap, not the token count. Output counts are often smaller than input counts. But output overtakes input cost whenever generated tokens exceed input divided by five. Moderate responses clear that bar easily. This dictates where to optimize. Cache any stable prefix first, since the ten percent rate makes it almost free. Then squeeze output length, because each saved output token is worth several input tokens. Trimming input below roughly a tenth of output is wasted effort. Watch agent loops, which multiply both terms across rounds and grow context every turn.