Developer proposes token cost as an observability metric

On 2026-11-13 21:26, an author on Juejin published an article arguing that token cost in AI applications should be treated as an observability metric. The piece begins by pointing out a practical problem: shortly after an AI application goes live, the first questions from leadership and finance teams are usually impossible to answer with confidence. For instance, when someone asks how much was spent this month, the only route today is to dig through cloud provider bills and estimate, and even after that it remains unclear which features absorbed the spend. A related question, the average cost per request, is equally hard to pin down with current setups.

The author's proposal is to stop relying on bill-by-bill guesswork and instead build token cost into observability tooling. Under this approach, teams would have a consistent, queryable basis for answering spending questions directly rather than reconstructing figures after the fact. The article also outlines a concrete method for breaking down cost by feature once an application is in production, explaining how to attribute overall token spend to individual functions so that the cost tied to each one can be checked.

As of now, this is the only publicly available write-up tied to the idea. The discussion rests at the stage of a proposal and a practical methodology, with no public signs of other organizations adopting the approach, building tooling around it, or weighing in on the proposal.