Pricing AI Products: Seats, Usage and Credits

Traditional software has very low marginal cost: serving one more user costs little, so per-seat pricing with high margins works well. AI changes this because each request consumes compute on GPUs, and the cost varies with prompt length, output length and which model is used. A light user and a heavy user on the same seat can differ in cost by a large multiple. Per-seat pricing is predictable for buyers and easy to sell, but the vendor carries the risk of heavy users. Vendors often add fair-use limits or rate caps to manage this. Usage-based pricing aligns cost with value and protects margins, which is why it dominates developer APIs. Its drawback is unpredictability: buyers struggle to budget, and some worry that every use costs money, which can suppress adoption. Credits sit between the two. A customer buys a fixed allowance, and different actions draw it down at different rates. Credits let a vendor price actions with very different costs under one balance and change the rates as model costs change. The downside is opacity: customers may not understand what a credit buys, and changes to credit rates can feel like hidden price increases. Hybrid models, such as a seat price with an included allowance and overage charges, are increasingly common because they combine budget predictability with protection against extreme usage. Changes in pricing can provoke strong reactions. Several AI coding tools revised plans during 2025 to limit or meter heavy use of expensive models, and some users objected that the changes were unclear. The lesson is that transparency about what is included, and advance notice of changes, matters as much as the structure itself. A common misconception is that usage pricing is always fairer; for many business buyers, predictability is worth paying for.