Blended cost

Because providers price input and output tokens differently, no single list price describes what a model costs you. Blended cost weights each rate by your workload ratio: a summarization pipeline might send 20 input tokens per output token, while a story generator might do the reverse, and the same model yields very different effective prices for each.

A useful formula is blended price = (input rate × input share) + (output rate × output share), computed from measured token logs rather than guesses. Reasoning models shift the blend heavily toward output because billed thinking tokens count as output.

Caching discounts, batch rates, and long-context surcharges all modify blended cost further, which is why realistic model comparisons start from telemetry on your actual traffic rather than from headline per-million-token rates.