Gartner says agent costs will rise fivefold by 2028 even as tokens get cheaper. Here is how to keep them in check

Falling token prices have led many enterprises to assume AI costs will fall with them. Agents are breaking that assumption.

Gartner forecasts that inference costs per agentic workflow will rise more than fivefold through 2028, even as token economics improve. The firm says that routing a task to an agentic reasoning model raises provider inference costs by at least 5X compared with a basic chatbot interaction, and often much more as task complexity grows. They call this the Inference Paradox. Better unit economics escalates the overall cost of AI without a clear path to commensurate and predictable value.

Gartner says improved efficiency is being reinvested in more powerful and more expensive models.

"In this way, token costs aren’t falling; they’re widening," Will Sommer, senior director analyst at Gartner, told VentureBeat.

Why cheaper tokens still mean bigger bills

There are two primary drivers behind the Inference Paradox.

The first issue is the model labs. Sommer noted that tokens are becoming more cost-efficient to generate. 

"However, instead of banking these cost-savings, foundational AI labs are investing in next-generation models that require more computational power, and thus cost more per token," he said.  

Older capabilities are getting much cheaper while frontier models are getting more expensive. Sommer pointed to Anthropic's Fable costing more per token than Opus as an example of the split.

The second driver is consumption. Greater token efficiency makes more complex workloads practical, and those workloads burn far more tokens.

"AI agents need 5-30x more tokens/tasks than basic AI chatbots for the same task," Sommer said.

Competitive pressure ties the two together. Enterprises that want to keep pace with peers adopt more advanced use cases, and those call for more capable and pricier models plus far more tokens than the previous generation of AI systems.

Costs will climb in jagged waves, not a straight line

Gartner does not expect a steady climb through 2028. The tension between constraining costs and remaining competitive with peers is pushing spend up and down, and spending has oscillated with new model releases.

"First-wave agentic capabilities led to a large increase in spending in the spring, driven in part by the tokenmaxxing fad," Sommer said. "We are currently in a belt-tightening phase."

Gartner expects several more waves as more capable reasoning agents, world models, and other AI technologies mature.

Enterprises are constraining costs unevenly. A September VB Pulse survey of 109 qualified respondents found that 21% of enterprises using an orchestration platform have nothing that automatically stops agents' token spending. Another 31% send routine, high-volume agent tasks to cheaper models to hold costs down. 

The fivefold estimate rests on Gartner's Tokenomics Forecast, an econometric model that predicts token costs over time from more than 300 inputs, according to Sommer. It accounts for both greater token-cost efficiency and sharply higher token consumption as AI is deployed against harder problems. Sommer said AI solutions that learn, or that operate across an enterprise or between multiple domains, need significantly more tokens to operationalize workflows. 

Five architectural changes to get ahead of the Inference Paradox

There are five key changes that organizations should consider to get ahead.

Invest in observability. AI FinOps is crucial to any cost-optimization effort. "Enterprises cannot optimize what they cannot observe," said Sommer.

Embed value-per-outcome in AI deployments. Require every AI feature to forecast against a wide range of uncertainty and track spend against a clear outcome metric. For example, tasks automated or cases successfully closed. Rapidly improve or deprecate low-performing workflows to protect margins.

Institutionalize inference tiering. Enforce automated routing of queries to the most cost-efficient model and block default invocation of frontier models for routine tasks. "Invest in agentic harnesses which define where, when and how models are deployed within autonomous systems," Sommer said.

Enforce minimum standards for AI execution. Set it and forget isn't going to work. Organizations need to define success thresholds, compliance overhead, and risk-mitigation costs upfront — and refuse to greenlight deployments until scenarios are stress-tested against token-price swings and compliance expenses.

Leverage proprietary data. Embed proprietary data fine-tuning and self-learning feedback loops to sustain differentiation and margins over time. "Efficient, domain-specific models generally outperform generic frontier models on reliability and tasks for specific workloads," Sommer said.