Seat-based pricing assumed that the main cost was building the software. Once it was built, adding another user was nearly free, so you charged per user. Usage-based pricing assumed compute was the main cost, so you charged per API call or per token. Both models map pricing to the underlying cost structure of the product.

AI-native products have a more complicated cost structure. The compute per task varies enormously depending on what the task is. A simple lookup costs almost nothing. A multi-step analysis across a hundred documents costs significantly more. Usage-based pricing by raw compute makes the customer’s bill unpredictable in a way that kills enterprise adoption. Seat-based pricing decouples the cost of the product from the value delivered, which makes it hard to expand revenue as customers get more value.

Credit-based pricing is the emerging answer. Customers buy a pool of credits — say, a thousand credits per month for a flat fee — and different actions within the product consume credits at different rates based on their actual compute cost. A quick lookup might cost one credit. A full document analysis might cost fifty. The customer has a predictable spend and a visible, manageable pool to work with.

This solves several problems at once. The product team can recost individual features as compute gets cheaper without changing the customer’s mental model. The customer can understand their usage and plan accordingly. Enterprise procurement gets a predictable line item. Sales gets a natural land-and-expand story: customers who hit their credit ceiling upgrade to larger pools.

The interesting design question is what to denominate credits in. The worst option is opaque units with no clear mapping to work done — this creates the “dark pattern” reputation that credit systems sometimes get. The best option is credits that map directly to a task the customer understands: one credit per document analyzed, ten credits per full deal package, and so on. When the credit unit maps to a unit of work the customer already thinks in, the pricing model becomes intuitive rather than confusing.

What’s happening in 2026 is that credit wallets are absorbing both seat-based and pure usage-based pricing for AI-native products. The seat model is too disconnected from value. The pure usage model is too unpredictable. Credits offer the middle ground: predictable for the buyer, usage-reflective for the seller, and flexible enough to accommodate widely varying task complexity in a single product.

The products that get this right build the credit denomination into the product from the start. The products that get it wrong bolt credits onto a seat-based product as an add-on and wonder why adoption is flat. +++