Gartner predicts that inference costs for agentic AI workflows will increase by more than five times by 2028. This trend persists despite a sharp decline in the cost of individual tokens, a phenomenon the firm describes as the Inference Paradox.
Analyst Will Sommer notes that agentic reasoning models increase inference costs by at least 5x compared to standard chatbot interactions. For complex tasks involving multiple reasoning steps and tool calls, these costs can reach 30x the price of a basic query.
Training expenses for these advanced models are also projected to rise by 2.5x. The overall increase in spending is driven by the deployment of more powerful models and a significant surge in token consumption required for multistep workflows.
To manage these costs, Gartner suggests three strategies: implementing inference tiering to route simple tasks to cheaper models, utilizing multimodal orchestration to coordinate different tools, and maintaining strict cost monitoring.
