Gartner: AI Agent Costs Will Rise Fivefold by 2028 Even as Token Prices Drop

Quick Facts

• Gartner predicts AI inference costs per agentic workflow will rise more than fivefold through 2028, even as token prices fall roughly 95% by 2030.
• AI agents require 5 to 30 times more tokens per task than basic chatbots, and over 40% of agentic AI projects will be canceled by end of 2027 due to cost overruns and unclear ROI.
• Gartner recommends five architectural changes, starting with observability and model routing, to keep agentic spending under control.

Cheaper AI tokens are not making enterprise AI cheaper. Gartner's new research shows that falling unit prices are being swamped by the volume and complexity of agentic workloads, sending total inference spend sharply higher.

The firm calls this the Inference Paradox. Gartner announced the findings on August 17, drawing on a research report by analysts Will Sommer and Sabine Zimmerhansl published in June 2026.

"Product leaders cannot rely on more efficient token economics to rationalize AI costs," said Sommer. "Each successive generation of AI capability will necessitate more, and often more expensive, tokens. There is no reliable, economical one-size-fits-all model on the horizon."

Why Agents Cost So Much More

The math is simple. Where a chatbot reads a query and responds, an agent must reason repeatedly, call external tools, and self-correct across multiple steps. According to Sommer, agents require 5 to 30 times more tokens per task than a basic chatbot handling the same job.

Gartner's own Tokenomics Model illustrates an eightfold cost spread across task types, from roughly $0.05 for a simple workflow to $0.40 for planning and learning tasks. Routing that planning task to a frontier reasoning model instead of a standard model raises the cost by a factor of five on Gartner's numbers, with extreme cases running 150 times the price of a basic chatbot call.

Three forces drive this dynamic. First, cheaper tokens make previously uneconomical applications viable, so companies build more of them. Second, AI labs are reinvesting cost savings into next-generation models that cost more per token, not less. Third, multistep agent workflows consume tokens at a pace simple applications never did.

Sommer pointed to Anthropic's model pricing as a real-world example. Claude Fable 5 is priced at $10 per million input tokens and $50 per million output tokens, double the input cost of Opus 4.8. Older model tiers get cheaper; frontier tiers get more expensive. "Token costs aren't falling," Sommer said. "They're widening."

Project Cancellations Are Coming

Gartner predicts more than 40% of agentic AI projects will be canceled by the end of 2027 because of escalating costs, unclear business value, or inadequate risk controls.

"Most agentic AI projects right now are early-stage experiments or proof of concepts that are mostly driven by hype," said Anushree Verma, senior director analyst at Gartner. "This can blind organizations to the real cost and complexity of deploying AI agents at scale."

Verma added that many use cases positioned as agentic today do not actually require agentic implementations, and that current models lack the maturity to autonomously achieve complex business goals at scale.

Despite the cost pressure, Gartner still sees wide adoption ahead. The firm predicts at least 15% of day-to-day work decisions will be made autonomously through agentic AI by 2028, up from 0% in 2024. It also forecasts that 33% of enterprise software applications will include agentic AI by 2028, up from less than 1% in 2024.

Five Ways to Control the Spend

Gartner outlines five architectural changes for technology leaders to contain costs as agent deployments scale.

• Invest in observability. "Enterprises cannot optimize what they cannot observe," Sommer said. AI FinOps tooling must come before any cost-optimization effort.
• Embed value-per-outcome tracking. Every AI feature should forecast against a range of outcomes and track spend against a clear metric, such as tasks automated or cases closed.
• Build agentic harnesses. Define where, when, and how models are deployed within autonomous systems rather than letting agents call frontier models by default.
• Use inference tiering. Route tasks to the smallest model capable of handling them. Sommer noted that domain-specific models generally outperform generic frontier models on reliability for targeted workloads.
• Deprecate underperforming features fast. Features that do not hit value targets should be cut quickly, not carried forward hoping performance improves.

"Defaulting to generic autonomous intelligence will result in unbounded costs orders of magnitude higher than those of optimized product ecosystems," Sommer said.

The warning is direct: falling token prices give executives a false sense of cost control. Without deliberate architecture and spending discipline, the bill for agentic AI will grow far faster than the value it delivers.