AI model providers may charge different rates for input tokens, output tokens and tokens served from a cache. A blended price combines those rates using an assumed workload mix, such as a chosen percentage of input, output and cached tokens, to make the pricing easier to compare.
The result depends on the mix. A model with a low input price but a high output price may look inexpensive for document classification and costly for long-form generation. A different share of cached requests can change the ranking again.
Blended price is an estimate, not a universal bill. It may omit tool-use charges, minimums, provider margins, retry requests, image or audio tokens and the extra calls required when a model fails. Self-hosted models have hardware and operating costs that are not naturally expressed as an API token price.
When a comparison presents one price per million tokens, look for the assumed input-to-output ratio, cache hit rate, currency and price date. Recalculate it using your own traffic before choosing a model.