Cost per Accepted Answer: A Method for Measuring the True Cost of LLM Inference
Year: 2026
Type: preprint
Abstract: Token price is the unit most teams use to reason about LLM cost, but it is not the unit they actually pay for. Teams pay for accepted answers, and every failed or retried attempt adds cost that token pricing hides. This preprint defines cost per accepted answer: a measurement method built on logical tasks, individually recorded attempts, acceptance recorded after review, and pricing snapshots that pin the exact rates used for every figure. It describes the design of an open source implementation, the AI Inference Control Plane, and a worked example with illustrative numbers. The method makes a cheaper model that needs more attempts directly comparable with an expensive model that answers cleanly, on the only number that matters: what one kept answer cost.
Author: Jaffar, Ahmed
Institution: Western University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225577656