What is token-based pricing?

Pricing an AI model by the tokens, small chunks of text, that it reads and writes, usually at different rates per model and for input and output.

How does it work?

Each request to a model is counted in input and output tokens and billed at the model’s rate. Anthropic’s S-1 calls its model “pay-for-intelligence”: usage metered in tokens and priced by model. Margin-Based Pricing for AI Products notes that output tokens cost three to five times input tokens. Reasoning models can cost ten to fifty times more than small models for the same task. An agent that loops through a document ten times costs ten times a single pass.

Why does it matter in finance ops?

The price per token is not the number a buyer needs. The S-1 brief puts it plainly. A cheaper token that needs three retries costs more.

Ask for cost per completed task. Token bills also arrive from many vendors on different meters. The bill moves with how hard people use the agents, so it behaves like an accrual, not a fixed subscription. Ramp now sells token spend management for this reason.

Some vendors take the token risk off the buyer. Sidetrade sells its AR agents by subscription with a volume of tasks and token equivalents, and no exposure to third-party per-token prices. The trade is that model choice moves to the vendor. For a company that sells AI products, token logs are only part of the cost. The margin essay records tool and data calls that can run at several times the model bill.

Where has it come up?

Which essays cover it?

Which terms are related?

Cite this

Couch, C. (2026). What is token-based pricing?. Ends at Cash. https://chriscouch.ai/what-is/token-based-pricing/

BibTeX

@misc{couch2026whatistokenbasedpricing,
  author = {Couch, Chris},
  title = {What is token-based pricing?},
  year = {2026},
  howpublished = {Ends at Cash},
  url = {https://chriscouch.ai/what-is/token-based-pricing/}
}