cost-modellisted

# AI Model Economics Turn “which model should we use?” into “what will it cost, can it handle the load, where is it officially available, and how long is the pricing basis valid?” ## Why use this skill Model pricing looks like a lookup table. In practice, three mistakes can make an estimate wrong by an order of magnitude: 1. **Treating token fees as project cost.** Human review can cost many times more than compute. 2. **Checking unit price but not rate limits.** A flagship model can cost more while offering only 1/60 of an older tier’s RPM. A workload may be affordable but unable to run. 3. **Quoting a retired SKU.** The price may be valid when estimated, but the model may be unavailable when the contract is signed. Use this order: **define the scope → assess all three axes → check whether the price is still valid → give the number.** ## Step 1: answer the three scope questions Before giving any number, answer these questions and state the answers in the output. Different scopes can produce estimates that differ by two orders of magnitude. ### 1. Is this compute cost or delivery cost? - **Compute cost** = tokens + tool calls + storage + subscriptions. This is the machine bill. - **Delivery cost** = compute + human review + fact checking + compliance review + localization + rework. This is the operating cost. Default to compute cost, **but explicitly list what is excluded**. In production content work, compute is often only a single-digit percentage of delivery cost