Price per million tokens

Virtually every AI API prices usage in dollars per million tokens, with distinct input and output rates. The convention makes list prices easy to scan, and it replaced the per-1K-token quotes common in earlier years as volumes grew.

Headline rates rarely tell the whole story. Providers layer on cached-input discounts, batch discounts, surcharges for very long contexts, and premium rates for priority capacity, and reasoning models consume billed thinking tokens that inflate effective output usage.

Because input and output are priced differently, two models can only be compared fairly through a blended cost that weights each rate by the input-output mix of your actual workload. A model with a cheaper input rate but pricier output can cost more overall for generation-heavy use.