If the unit cost of running AI models falls, shouldn't the cost of using AI agents fall too? Viewed that way, the outlook Gartner presented on August 17 seems to run in the opposite direction. The firm forecasts that AI inference costs per agentic workflow will more than quintuple by 2028. This isn't a company's total AI spending or what gets billed to customers—it's the provider-side inference cost per workflow that product owners manage.
The driving force behind this is a shift toward using more powerful, expensive models through longer processes, made possible by access to cheaper inference. An agent doesn't necessarily return a single answer and stop. It plans, calls tools, evaluates results, retries if necessary, and hands off work to other agents. Rather than unit price, it's the number of calls, the context passed along, the model chosen, and the conditions for termination that determine the cost of a single task.
Gartner calls this phenomenon the "Inference Paradox"—the idea that improvements in cost efficiency expand the use of more capable models and drive up total costs, while value doesn't necessarily increase at the same predictable rate. For enterprise management, this means tracking not just provider inference costs, but the "cost per outcome"—the total cost of bringing a single task to an acceptable result.
What exactly does "over 5x" refer to?
According to Gartner, compared to a basic chatbot, routing tasks to agentic reasoning models results in provider-side inference costs that are at least 5 times higher. The gap can widen further as tasks become more complex. The "over 5x by 2028" figure is likewise a per-workflow forecast—it doesn't mean that billing will uniformly increase fivefold across all companies, use cases, and agents.
The public announcement doesn't provide absolute dollar figures for 2026 or 2028, industry breakdowns, the estimation model, or sample data. As a result, this multiplier cannot be converted into a dollar amount. It's also a different metric from the inference cost borne by model providers, the API fees paid by customers, or the total cost including internal labor within a company.
The pathway to ballooning costs can't be explained by token unit price alone. Agents chain together reasoning to decide the next step, calls to external tools, and verification after reading that output. Using multiple agents adds more role-specific responses and handoffs. Compared to a short chat response, if the design calls for more tokens and inference cycles per workflow, the effect of falling unit prices gets absorbed along the way.
Why this doesn't contradict predictions of falling unit costs
In March, Gartner had forecast that costs for generative AI providers running inference on trillion-parameter large language models would fall by more than 90% by 2030 compared to 2025. Compared to similarly sized early models from 2022, LLMs in 2030 are expected to be up to 100 times more cost-efficient. This is a forecast about falling provider-side costs for running models of comparable scale.
Meanwhile, the company expects agentic models to require 5 to 30 times more tokens per task than standard generative AI chatbots. This 5-to-30x figure refers to token consumption—not a multiplier on a company's total costs or customer billing. The 90%-plus decline is a provider cost outlook, not a direct forecast of customer-facing API pricing.
The two figures have different denominators and different scopes—they can't be multiplied together to calculate future customer costs. The year, model scale, token type, provider margins, and task complexity aren't aligned between them. Gartner itself doesn't expect per-token cost reductions to be fully passed through to enterprise customers.
Falling unit prices don't necessarily suppress usage—they can also lower the barrier to experimenting with more advanced reasoning. The company's advice—to delegate routine processing to cheaper models while reserving frontier models for complex, high-margin reasoning tasks—is less about reducing overall usage and more about re-selecting which capability to use for each job.
In implementation, parallelization amplifies token usage
Anthropic's explanation of Claude Research, published in June 2025, offers one example of where this amplification occurs. According to the company's internal data, agents used about 4 times more tokens than a chat interaction, while multi-agent systems used about 15 times more. This figure is specific to Anthropic's research system—it's not a market-wide average, nor direct verification of Gartner's forecast.
The company reports that a configuration with Claude Opus 4 as the lead agent and Claude Sonnet 4 as sub-agents outperformed a single-agent Claude Opus 4 by 90.2% in internal research evaluations. It states that 95% of the variance in evaluation results can be explained by three factors—token usage, tool calls, and model choice—with token usage alone accounting for 80%. However, this too is an internal evaluation. It doesn't follow that increasing tokens always improves cost-effectiveness.
Anthropic explains that multi-agent setups are well-suited to high-value tasks involving substantial parallelizable work, work that doesn't fit within a single context window, and complex tool usage. They're less suited to tasks with heavy shared context or dependencies, and even coding tasks sometimes offer limited room for parallelization. The decision to add more roles shouldn't hinge on the number of agents itself, but on whether the gains from splitting a task justify the additional inference costs incurred.
Managing by outcome, not by request
Gartner states that in the shift from basic chatbots to autonomous agentic AI, financial management needs to change from "cost per request" to "cost per outcome." Even with a cheap model, if failures lead to more retries or human corrections, the total cost of reaching an accepted result won't necessarily fall. Conversely, even expensive inference can be the right call if it produces an appropriate outcome on the first attempt.
As countermeasures, Gartner points to differentiating between reasoning tiers, routing, and orchestration. Applying general-purpose autonomous intelligence by default to every task can end up costing orders of magnitude more than an optimized suite of tools. A design that routes routine, high-frequency processing to small or domain-specific models while reserving top-tier models for genuinely difficult reasoning makes both cost and performance comparable.
AWS's Agentic AI Lens also points out that the plan-execute-verify-reflect loop, along with coordination among multiple agents, amplifies costs. It recommends setting explicit termination conditions, capping the number of iterations, establishing token budgets, and minimizing the context passed along in handoffs. Keeping hierarchies shallow and measuring inference costs separately from coordination costs makes it possible to trace which steps are racking up cycles without contributing to the outcome. That said, this is design guidance—not experimental results demonstrating the cost-reduction rate of each method.
Measuring between forecasts of adoption and forecasts of cancellation
In June 2025, Gartner predicted that more than 40% of agentic AI projects would be canceled by the end of 2027 due to rising costs, unclear business value, or inadequate risk management. At the same time, it also forecasts that by 2028, agentic AI will autonomously handle at least 15% of day-to-day operational decisions, and that 33% of enterprise software will include agentic AI. Widespread adoption and individual projects being cost-effective are two separate matters.
What companies need to verify isn't whether they've lowered the per-token price for each model. It's whether they can track, workflow by workflow, the cost per accepted outcome, the success rate, the number of retries, and the burden of human review. A design that increases inference only becomes a viable operational choice once it's confirmed that the resulting outcomes justify the additional cost.