AI Architecture Cost Model and Checklist

1. Cost model

Enter your traffic and token counts, and pick a model to load its published list price, or type your own prices. Leave a field empty to treat it as zero. This estimates token cost plus the other monthly costs you list. It does not include taxes, volume discounts, cache-write or storage fees, network egress or engineering time.

Fill with an example Example traffic and costs. The model prices are the published list prices from 27 October 2026.

Calculate

2. Architecture decision checklist

Forty questions an architecture or security review is likely to ask of an AI system. Tick an item when you have a written answer, not when you have an opinion. Your ticks are kept in this browser.

Download open decisions as CSV Print or save as PDF Clear all ticks

How the cost is worked out

• Monthly requests = requests per day x days per month. Model calls = requests x calls per request.
• Input tokens per call = prompt tokens + retrieved context tokens. Output tokens per call are as entered.
• Main model cost per call = input tokens x blended input price + output tokens x output price. The blended input price = (1 - cache share) x input price + cache share x cached price. If you leave the cached price empty, caching is ignored.
• If you set a cheaper model, it takes its share of calls at its own prices, with no caching assumed. The main model takes the rest.
• Total = main model + cheaper model + the other monthly costs you list. Cost per request = total divided by monthly requests.
• Model prices are published list prices in US dollars. They are converted to your currency at the rate shown. Cache-write fees and storage fees are not modelled.
• Levers: cache share up 20 points (needs a cached price), 30 more points of calls sent to the cheaper model (needs one), retrieved context cut by 30%, output tokens cut by 20%.