After OpenAI made per-token pricing the default mental model for AI capabilities, every SaaS team adding AI features started asking: should we charge per token? The answer is almost certainly no, and the reasoning matters.
Token pricing is a compute pricing model. It makes sense for OpenAI because OpenAI is, at its core, a compute business. The model runs inference, inference costs tokens, the pricing reflects that cost. Clean.
For a SaaS business, you're not selling compute. You're selling a business outcome — a deal closed, a contract reviewed, a customer retained, a decision made. The number of tokens required to produce that outcome is an implementation detail that's invisible to and irrelevant to your customer.
When SaaS companies adopt token pricing for their AI features, they create three problems:
Metric opacity. Most users have no intuition for what a token costs or how many tokens a workflow uses. A pricing model built on an opaque metric creates buyer anxiety, not buyer confidence.
Misaligned incentives. If you charge per token, you're incentivized to use more tokens per outcome. If you charge per outcome, you're incentivized to do it efficiently. The first incentive is the wrong one.
Margin vulnerability. Your cost per token is set by the model provider and will decrease over time as models improve. If your pricing is also per token, your revenue decreases with your cost — you get none of the margin benefit of model improvement.
The right approach: charge on a business value metric that your customer understands and cares about. Define the outcome (documents processed, workflows completed, insights generated). Price on the outcome. Use tokens internally as a cost input, not as a customer-facing pricing unit.
Price for outcomes. Optimize tokens internally. Your customers will thank you and your CFO will too.