enterprises deploying AI agents should brace for a cost surge that may surprise even the most optimistic budget planners. Gartner’s latest forecast projects that agent operational costs will rise fivefold by 2028, even as underlying token prices continue their downward trend.
The paradox stems from how agents consume compute. While language model API calls have become dramatically cheaper—with Anthropic and OpenAI cutting prices multiple times in 2026—agents multiply those calls across loops, tool invocations, and retry logic. A single agentic workflow might execute dozens of model calls where a simple chat uses one.
Why costs scale differently. Traditional LLM applications pass a prompt and receive a completion. Agents, by contrast, operate in continuous loops: observe state, reason about actions, execute tools, observe results, and repeat. Each iteration compounds inference costs. Add multi-agent architectures where specialized agents collaborate, and the cost multiplier grows exponentially.
Gartner analysts note that enterprise agent deployments often run continuously in production, handling thousands of concurrent requests. The compute required to maintain agent state, manage context windows, and orchestrate tool calls dwarfs the baseline LLM API cost.
The mitigation path. The forecast isn’t all grim. Gartner suggests several strategies: optimize agent architecture to reduce reasoning cycles, implement aggressive caching of intermediate results, and select models that match task complexity rather than defaulting to frontier models.
Early adopters report wide variance in agent costs depending on implementation quality. A well-designed agentic customer service system might cost $0.02 per interaction, while a poorly optimized research agent burning through context could run $2.00 or more per task.
The projection adds urgency to the agent cost management conversation. As enterprises move from pilots to production agent deployments in 2027, those who haven’t addressed cost architecture may face budget shocks.