Cost per task
Cost per task is the total spent to complete one unit of work from start to finish, every request, retry and delegated call included, as opposed to the price of a token or of a single request.
Price lists quote tokens; what you pay for is finished work. An agent completes a task through many requests (exploring, editing, running tests, correcting itself), and how many it needs varies with the model and the effort level. The model that is cheapest per token is therefore not necessarily the cheapest per task. Anthropic's own cost guidance, as of September 2026, recommends comparing models on cost per completed task, noting that a more capable model often finishes with fewer turns, less searching and less backtracking.
Failures belong in the number. A failed attempt still bills its tokens, and so does the retry, so a model that is cheap per request but fails more often can cost more per task. Rework counts too: an answer that has to be corrected in a second session was not cheap.
Token counts also stop being comparable across tokenizers. Per Anthropic, the same text is about 30% more tokens on its 4.7-and-later models than on earlier ones, so a comparison of tokens used per task penalizes the newer models by construction, while a comparison of prices per token ignores that each of their tokens covers less text. Dollars per completed task survive the change.
Measuring it takes a defined task, a check that says whether it succeeded, and every cost counted: the main session, subagents, thinking tokens and retries. On a subscription, where no per-token bill exists, the total is usually expressed as an API-equivalent cost.