Blended Rate Calculator.

Why list price is the wrong number

A fleet almost never runs at list price and never runs on one model. Vendors quietly route a share of your traffic to cheaper models, you route some yourself, and both moves happen without appearing on an invoice line. The result is that the rate in your budget is a number nobody can reproduce, which is exactly the failure mode that makes a forecast wrong before anyone notices.

Two things fix it. First, price the blended rate, weighted by actual traffic share, and re-derive it monthly from provider exports rather than from a procurement sheet. Second, divide by a denominator that means something outside the engineering team, so the number survives a budget conversation instead of a code review.

The unverified alias input is the honest part of the model. If part of your traffic is served by a model you cannot prove from the response, the true spend sits somewhere between two bounds rather than on a single figure. The band below is that uncertainty, not a confidence interval. Pinning down which model actually serves a request is worth more than any discount negotiation you can run afterwards.

Method

Blended input rate is the traffic-weighted average of your tiers. Cache reads are billed at 10% of list input, a standard long-TTL cache-read discount, so cached input tokens are discounted before the blend is applied. The band assumes unverified traffic resolves entirely to the cheaper tier at the low end and entirely to the premium tier at the high end. Cost per outcome is total monthly spend divided by the outcomes you declared. Change the denominator and you change the number, which is why the definition matters more than the figure.