1. Gartner Warns That AI Inference Costs Will Climb Even as Tokens Get Cheaper
Research firm Gartner said in a forecast published on August 17, 2026, that AI inference costs per agentic workflow will rise by more than fivefold through 2028, even though the price of each individual token is falling sharply. The firm calls the pattern the Inference Paradox and cautions that cheaper units do not mean smaller bills.
The warning lands on a fundamental assumption in many corporate AI budgets: that the steady decline in the price of running models will make every new project cheaper than the last. Gartner argues the opposite is likely to hold once software moves from simple chat exchanges to autonomous, multistep agents. Executives who built their financial plans around falling token prices, the firm suggests, should return to those numbers and test them again.
Two terms are worth defining for readers outside the field. Inference is the stage at which a trained model produces an answer, and every answer consumes computing power that someone has to pay for. A token is the small chunk of text, roughly a word or part of a word, in which model providers measure and price that work. An agentic workflow is a task in which a model plans, calls software tools, checks its own results and repeats the cycle until a goal is met, rather than replying once to a single prompt.
2. What did Gartner forecast about AI inference costs?
Gartner expects the inference cost of completing an agentic workflow to grow by more than five times through 2028. That projection sits alongside a steep decline in the price of each token charged for foundation models, so the two trends run in opposite directions at the same time.
The distinction between unit economics and total spend is the heart of the forecast. A falling price per token describes how cheap one small piece of work has become. The size of an invoice, however, depends on how many pieces of work a task requires, and agent-style software needs far more of them than a conventional chatbot reply does. The result, in Gartner's framing, is that the cost of completing real work keeps rising even while the cost of each token keeps dropping.
Gartner analyst Will Sommer put a floor under the comparison. He said that routing a task to an agentic reasoning model raises the provider's inference costs by at least five times compared with a standard chatbot exchange. The five-times figure is a minimum rather than a typical or maximum case, and it is the foundation on which the broader fivefold projection rests.
3. How much more can an agentic task cost?
The gap widens as tasks become more elaborate. Gartner indicates that additional reasoning steps, repeated tool calls and more complicated workflows can push the cost of a single task to between five and thirty times that of a basic chatbot interaction. A workflow that looks inexpensive in a demonstration can therefore become a substantial recurring expense once it runs at scale inside a business.
Training is part of the picture as well. Gartner estimates that the more capable models behind these agents cost roughly 2.5 times as much to train. That means the added expense is not confined to the moment a customer uses a product: it is also built into the process of creating the models that power it, and providers have to recover those outlays in some form.
Taken together, the figures describe a cost structure in which the headline price list and the final bill can drift far apart. A team that estimates spending by multiplying expected tokens by a published rate may undercount badly if its software ends up reasoning for longer, calling more tools or selecting a larger model than planned.
4. Why do cheaper tokens still produce bigger bills?
Because the number of tokens consumed per task is growing faster than the price per token is shrinking. Gartner points to three forces that combine to produce this outcome, and each one reinforces the others.
The first is improving model economics. When foundation models become cheaper to run, applications that were previously too expensive to justify become viable, and companies respond by building more of them. Lower prices widen the range of projects that clear a budget threshold, which raises overall consumption rather than holding it flat.
The second is the use of more powerful models. Complex applications tend to rely on more capable and more expensive systems, because teams reach for the heavyweight option when a task is hard. A cheaper entry-level model does little to reduce spending if the product's quality depends on the premium tier.
The third is token consumption itself. Workflows that unfold across many steps cause a pronounced jump in usage, since each stage can involve fresh reasoning, a tool call and a review of earlier output. Multiplied across thousands or millions of tasks, that behavior turns a small per-token charge into a large aggregate cost.
5. How does the forecast square with earlier price projections?
The August forecast does not contradict Gartner's earlier view on token pricing so much as qualify it. In a previous 2026 projection, the firm had expected the cost of tokens for large models to fall by more than 90 percent by 2030. The new analysis accepts that a decline of that scale is plausible, and then argues that it will be overwhelmed by changes in how the technology is used.
The reasoning is that a price drop above 90 percent is offset by workflows demanding far more tokens, and by applications that start from expensive models in the first place. The cost problem therefore tends to correct itself in the wrong direction: as one input becomes cheaper, buyers use more of it and trade up to better versions, and the total bill grows.
That dynamic is familiar from other technology markets, where lower unit prices have often expanded demand enough to raise total expenditure. What makes the AI case notable is the speed at which a single task can multiply its own consumption through repeated reasoning, a feature that traditional software workloads rarely share.
6. Why the minimum matters for planning
Budget owners tend to anchor on averages, so Gartner's emphasis on a floor deserves attention. By describing five times as the least that a reasoning-based task adds to provider costs, the analysis tells planners that the downside scenario is the more realistic place to start, and that any improvement over it is a bonus rather than a baseline.
This also changes how vendors and customers negotiate. A provider that absorbs inference costs inside a flat subscription carries the risk that heavy agent usage erodes its margins, while a customer billed per task carries the risk of an invoice that outruns expectations. Either arrangement benefits from a clear estimate of how many reasoning steps and tool calls a typical job requires, which is information many teams do not yet collect.
For that reason the forecast is best read as a prompt to measure rather than as a prediction to accept passively. The firm's own recommendations all begin with knowing where tokens are going, and they assume that the answer will differ from what a simple price list suggests.
7. What does the Inference Paradox mean for companies budgeting for AI?
For businesses, the principal consequence is that return on investment becomes harder to predict. Gartner presents the paradox as a challenge for product leaders and for anyone accountable for the financial results of AI programs, because better unit economics do not guarantee lower total costs.
A product manager who forecasts margins on the assumption that model prices will keep falling could find that the same product, once it gains reasoning ability and tool access, costs several times more per completed task. Pricing decisions made to customers on the earlier assumption would then squeeze profitability. Internal projects face a parallel risk: a pilot that appears affordable can become expensive when promoted to wide use.
The uncertainty is also a governance matter. Because cost can vary by a factor of five to thirty depending on how a workflow is designed, finance and engineering teams need a shared understanding of which tasks run on which models. Without it, spending can grow quietly in the gap between what was budgeted and what the software actually does.
8. What should product leaders do about rising AI agent costs?
Gartner recommends three practical approaches. The first is inference tiering, which means directing simple tasks to cheaper models and reserving the most expensive systems for work that genuinely demands them. A routine classification or lookup does not need the same model as a complex multistep analysis, and sorting requests this way keeps premium spending targeted.
The second is multimodel orchestration, the coordination of several models so that no single costly system handles everything. In this arrangement a lightweight model can handle the first pass, a stronger one can take over when difficulty rises, and specialized models can be used where they perform best. Orchestration treats model choice as an ongoing operational decision rather than a one-time purchase.
The third is vigilant cost monitoring. Since a single workflow can cost between five and thirty times what a chatbot query does, Gartner's approach calls for tracking spending closely at the level of individual workflows, not only in monthly totals. Visibility makes it possible to spot an expensive pattern early, adjust the routing and keep AI investments tied to measurable outcomes.
The broader message from the forecast is one of caution rather than alarm. Falling token prices remain a real and useful trend, but Gartner's analysis indicates that they are not a reliable guide to what a company will pay as its AI systems grow more capable and more autonomous. Organizations that plan for the total cost of finishing a task, and not just the price of a token, will be better placed to judge whether an agent is worth running.