Don’t Price AI Applications Per Token

The a16z opinion piece argues that token-based pricing—appropriate for model providers—often misaligns incentives when carried into AI applications. Instead, vendors should price at the highest layer of measurable customer value: tokens for raw model access, credits that map to recognizable work for variable application tasks, and outcome-based pricing when business results are observable and attributable. Well-designed credit systems should abstract infrastructure complexity, explain relative effort, and preserve commercial flexibility. The article recommends hybrid approaches (seats, credits, and token pass-through for volatile model costs) and cites Clay’s 2026 pricing memo as an example of separating Data Credits from Actions and selectively passing through expensive model costs.