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 these more capable models at an estimated 2.5x higher. So the expense shows up on both ends: building the models and running them.
The three drivers
Gartner identifies three forces behind the cost anomaly.
The first is improving foundational model economics. Cheaper tokens make it viable to build applications that would have been prohibitively expensive before, so companies build them.
The second is the deployment of more powerful, and therefore pricier, models for complex applications. When the job is hard, teams reach for the heavyweight option.
The third is a pronounced jump in token consumption from multistep agent workflows.
How we got here
This forecast sits alongside an earlier Gartner projection from 2026 that the cost of large model tokens would fall by over 90% by 2030. On paper, that sounds like the industry’s cost problem solving itself.
The new forecast argues it will not. A 90%-plus drop in per-token pricing gets swallowed when workflows need dramatically more tokens and lean on more expensive models to begin with.
What this means for companies building on AI
Gartner frames the paradox as a direct challenge for product leaders and anyone responsible for capturing a return on AI investment. Better unit economics do not guarantee lower overall costs, and the path to ROI becomes harder to predict.
The firm’s recommended playbook leans on three approaches. The first is inference tiering, which means routing simple tasks to cheaper models and saving the expensive ones for work that actually needs them.
The second is multimodal orchestration, coordinating different models and modes so that no single costly model does everything. The third is vigilant cost monitoring, which becomes essential when a single workflow can range from 5x to 30x the cost of a chatbot query.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.