i
News
All Research News
News · 2026-10-16

Gartner sees AI-agent costs rising fivefold by 2028

@neuronium_ai @neuronium_ai

Gartner expects AI-agent costs to rise fivefold by 2028, even as token prices fall. The forecast, attributed to Gartner senior director analyst Will Sommer, points to a mismatch: older AI capabilities are getting cheaper, while frontier models and the workloads built around them are becoming more expensive. For companies adopting agents, the question is not simply what a token costs, but how many tokens a task consumes and whether the result justifies the bill.

Cover: Gartner sees AI-agent costs rising fivefold by 2028

Why cheaper tokens can mean higher bills

Sommer describes two forces behind the inference paradox. Model developers use efficiency gains to build more capable systems that require more computing power, rather than passing all of the savings on to customers. At the same time, cheaper and more efficient token use makes complex tasks practical, increasing consumption.

The gap between older and more advanced models is widening. Sommer cited Anthropic Fable as an example, saying its tokens cost more than Opus tokens. And agents can use 5–30 times as many tokens per task as ordinary chatbots.

Competition adds pressure. Companies trying to keep pace with rivals adopt more ambitious workflows, which can require both more capable, costlier models and far more tokens than earlier AI systems.

Spending will come in waves

Gartner does not expect costs to climb steadily through 2028. Spending fluctuates as new models arrive and companies weigh cost control against competitive pressure. Sommer said the first wave of agent capabilities drove a sharp increase in spending in the spring, partly because companies were eager to use as many tokens as possible. They are now shifting toward savings.

Gartner expects further spending waves as more capable reasoning models, world models and other AI technologies develop.

Companies are already taking different approaches. In a September VB Pulse survey of 109 qualified respondents, 21% of companies using an orchestration platform had no automated mechanism to stop agents’ token spending. Another 31% routed routine, high-volume agent tasks to cheaper models.

The fivefold forecast comes from Gartner’s Tokenomics Forecast, an econometric model that, Sommer said, tracks token costs over time using more than 300 inputs. It accounts for both improved token efficiency and rising consumption as AI takes on harder tasks. Systems that learn or operate across an entire company or multiple domains need substantially more tokens to run their workflows.

Five ways to put a ceiling on costs

Gartner recommends five architectural changes to manage the trade-off:

1Make usage visible. AI FinOps—the financial management of AI spending—starts with observability. Companies cannot optimize costs they cannot see.
2Tie spending to outcomes. Forecast costs for each AI function across a wide range of uncertainty, then compare them with a clear measure of value, such as tasks automated or support cases resolved. Improve or shut down workflows that perform poorly.
3Route tasks by inference tier. Send requests automatically to the least expensive model that can handle them, rather than defaulting to frontier models for routine work. Sommer also recommends investing in agent frameworks that determine where, when and how models are used in autonomous systems.
4Set minimum performance standards. Define success thresholds, compliance costs and risk controls before deployment. Stress-test systems against changing token prices and compliance costs; a set-it-and-forget-it approach will not work.
5Use proprietary data. Build in fine-tuning on company data and self-learning feedback loops. Sommer said efficient, domain-specific models are usually more reliable than general-purpose frontier models on specialized tasks and can help protect differentiation and margins.

I think the forecast’s most useful signal is not the fivefold figure itself, but the cost-control gap: more than one in five surveyed companies using orchestration had no automatic stop for agent token spending. As workloads grow more ambitious, choosing a cheaper model matters less if no one is tracking when an agent should stop.

Source

Daily AI news

Every day we pick what actually matters in AI and explain it plainly — no hype, no filler. Subscribe if you want to follow where the industry is going.

Only what matters — every day

Follow on X