- Published on
Why I tried to kill token billing (and why we kept it)
- Authors
- Name
- DeepHorus
Metronome tried moving beyond token billing as a customer-facing price but retained it as internal infrastructure. The company says businesses still need token-level metering to track usage, manage markups and protect unit economics. Most do not want those costs to define their product pricing.
Token billing charges customers for consumed model tokens, plus a markup on provider costs. It is easy to explain and helps guard against runaway compute expenses. For model providers, tokens are the product; for most other AI companies, tokens and subscription seats do not reflect customer value.
Passthrough pricing can make a product resemble a commodity markup. An illustrative invoice can show usage across 10 models, with markups of 10%, 5% and 25%. As models become cheaper and more interchangeable, customers may challenge markups or bypass the provider, potentially pushing margins toward zero.
Metronome argues that invoices should express a provider’s value rather than reveal its model mix, input costs and margins. It is highlighting unified credits as a better customer-facing model. Customers would spend a common balance as operations consume credits at rates based on underlying token use, creating a credit-based burndown.
An invoice could identify dataset enrichments or generated images instead of enumerating models and markups. Complexity remains behind the scenes, where customers can route models upstream and tune margins as workloads change. Token billing continues to provide cost tracking, while credits present consumption in terms of work performed.
Metronome describes unified credits as a stepping stone toward output-based pricing, which it views as closely aligned with value today. The emerging model separates internal controls from external billing: token data supports routing and margin decisions, while customers pay for understandable outputs. It aims to preserve pricing flexibility without exposing the AI cost stack.
