An important paradox in the AI industry: the price of individual AI tokens is falling rapidly, but the total cost of running sophisticated AI agents is rising. Gartner predicts that token costs could fall by about 95% by 2030, while the inference cost associated with agentic workflows could increase more than fivefold through 2028. The reason is that agents do much more than answer a single prompt—they reason through problems, make plans, re-evaluate their work, use tools and sometimes communicate with other agents.
A traditional chatbot might process a question and generate an answer in a relatively short interaction. An AI agent, by contrast, may perform dozens or hundreds of model calls while completing a task. It might break a problem into smaller jobs, inspect information, call external tools, check its own results and retry when something goes wrong. As models become more capable, developers are also asking them to perform more complex reasoning. Consequently, a cheaper token does not necessarily translate into a cheaper completed task.
Gartner describes this as an “inference paradox” and warns about what it calls a “token-deflation illusion.” Companies may see falling prices from AI providers and assume that their overall AI budgets will fall accordingly. But if cheaper inference encourages much heavier usage—and if each task requires increasingly sophisticated reasoning—the total workload can expand faster than unit costs decline. The economic question is therefore shifting from “How much does a token cost?” to “How much does it cost to complete a useful business outcome?”
The broader implication is that enterprises will need to become much more disciplined about agent architecture, model selection and workload economics. Simply choosing the cheapest model will not solve the problem if an agent uses it excessively or repeatedly. Companies may instead need to route simple tasks to lightweight models, reserve expensive reasoning models for difficult decisions, limit unnecessary agent loops and continuously measure the cost of completed workflows. The article's central message is clear: AI is becoming cheaper at the unit level while becoming more expensive at the work level—and the winners will be organizations that optimize for the cost of the outcome, not just the price of inference.
