Cost per task

Price per token tells you little about what an agent actually costs, because one task can involve dozens of model calls, growing context, retries, and paid tool calls. Cost per task adds all of that up for one finished outcome: one resolved ticket, one research report, one merged pull request.

It is the number that matters for decisions. A more capable model with higher per-token prices can be cheaper per task if it finishes in fewer steps and fails less often. Compare cost per successful task, not just cost per attempt.

To measure it, log token usage for every call in a trace and roll it up per task. The common levers for reducing it are prompt caching, trimming tool outputs, routing simple steps to smaller models, capping steps, and batching work that is not urgent.