Gartner forecasts fivefold rise in AI agent costs by 2028

AI is getting cheaper to run. AI is also getting more expensive to run. According to Gartner, both statements are true at the same time. In a forecast released on August 17, 2026, the research firm predicted that inference costs per agentic workflow will rise by more than fivefold through 2028. That is happening even as the price of each individual token falls sharply. Gartner calls this the “Inference Paradox.” Executives who budgeted around falling token prices may want to reread their spreadsheets. Cheaper units, bigger bills The core of the forecast is a split between unit economics and total spend. Per-token costs for foundational models keep dropping. The total cost of getting real work done with those models keeps rising. Gartner analyst Will Sommer pointed out that agentic reasoning models push provider inference costs up by at least five times compared with a standard chatbot exchange. That multiple is the floor, not the ceiling. As tasks get more involved, the costs stack up. More reasoning steps, more tool calls and more elaborate workflows can drive costs to between 5x and 30x what a basic chatbot interaction would cost. Gartner’s figures also put training costs for the...