A new industry analysis breaks down the real economics of AI agents — from $5-per-million-token frontier model costs to banks spending up to $200,000 on multiagent workflows, and how to cut those costs dramatically.
Companies rushing to adopt agentic AI are discovering that the cost math is far more complicated — and at times far more expensive — than the headline price of a token suggests. According to a new industry analysis published in August 2026, individual token costs keep falling, but overall enterprise AI spending is spiking anyway, driven by factors most leaders aren’t tracking closely enough.
GPT-5.5 Costs Roughly 24x More Per Output Token Than an Earlier Lightweight Model
As of early July 2026, running a frontier model such as OpenAI’s GPT-5.5 costs about $5 per one million input tokens and $30 per one million output tokens. Compare that with an earlier, lightweight model at just $0.20 and $1.25 for the same volumes, respectively — nearly a 24x jump in output-token cost for the more capable model. The analysis’s authors stress that different models offer genuinely different capability levels, but note that many organizations still default to the most expensive model option even when a task doesn’t require it — a habit with real cost consequences at scale.
Some Banks Are Spending $100,000–$200,000 to Run a Single Multiagent Team
The scale of the problem becomes clear in real deployment numbers. Customer-facing AI agents at some banks can cost $20,000 to $30,000 to run a single-agent workflow, and $100,000 to $200,000 to run a full multiagent team, according to the analysis. Multiply that by the number of times a multiagent team runs, and by the number of workflows a company tries to automate this way, and costs can spiral quickly without careful oversight.
Token Costs Are Often Just 20–25% of the Real Bill — Human Oversight Eats the Rest
One of the analysis’s clearer findings cuts against conventional wisdom: for an AI agent handling a customer service task in banking, token costs frequently represent just 20 to 25% of the total variable run cost. Human oversight — typically performed by functional and risk experts — accounts for 70 to 75% of variable costs instead. In banking customer onboarding specifically, the analysis expects 10 to 20% of agentic runs to require review by risk and functional experts. The takeaway: most companies optimizing hard on model selection and token costs are optimizing the smaller line item, while the bigger cost driver — how much human review a workflow requires — often goes unaddressed.
One Bank Onboarding Workflow Needs 5 to 7 AI Agents and Up to 4 Human Teams
The analysis walks through a real-world example: a customer opening a bank account. Completing that single task end-to-end requires five to seven separate AI agents, multiple deterministic (rules-based) systems, and two to four teams of people providing oversight — a far more complex picture than the “one agent, one task” mental model many leaders start with.
Doubling Customer Volume Barely Moved the Cost Needle in One Real Example
Scale economics can work strongly in a company’s favor once the fixed costs of building an agent are in place. In one cited example, using a conversational agent to onboard 2,500 new customers a year costs $10,000 to $15,000. Doubling that volume of new customers pushed costs up to just $15,000 to $20,000 — nowhere near double, because the fixed infrastructure and orchestration costs get spread across far more transactions.
The Real Win: Cutting Onboarding Cost From $150 to as Low as $10 Per Customer
Using standard benchmarks, the analysis found that the fully loaded cost to complete an onboarding workflow — accounting for every agent, system, and human reviewer involved — can fall from roughly $50–$150 per customer down to about $10–$30 per customer once a workflow is properly designed and optimized. That’s the metric the analysis argues actually matters: not what a single agent costs to run, but the fully loaded cost to complete real, finished work, measured against the value it generates.
The Discipline Leaders Need: Budget for Change, Not Just for Launch
Because agent economics shift constantly as models improve and pricing changes, the analysis argues production agents often need adjustments every couple of days as new foundation models, protocols, and business requirements emerge — sometimes to the point that it makes more sense to retire an agent than keep tweaking it. The recommendation: build a dedicated operations discipline for managing agents over time — similar to how companies built “FinOps” to manage cloud spending — and incorporate AI unit economics directly into quarterly business reviews, so leaders can see which workflows still justify their spend and which need to be redesigned, consolidated, or shifted to cheaper models.
The Questions Leaders Will Have to Answer Next
The analysis closes by flagging a set of second-order questions companies are only beginning to confront: when AI costs should be passed on to customers versus absorbed internally, how much pricing flexibility to build into budgets, and how existing cost-allocation and governance practices need to evolve as AI agents increasingly work across department boundaries. As the authors put it, this is still “the foothills” of the AI economics learning curve.
Disclaimer: The cost figures cited in this article (token pricing, agent run costs, ROI ranges) are drawn from third-party industry analysis and real-world examples cited by the source; actual costs will vary by organization, vendor, and deployment. This is business commentary, not financial or investment advice.
The statistics, projections (including the ~$0.8 trillion value estimate), and case studies cited in this article come from third-party survey research and company examples; individual organizational results will vary. This is business commentary, not financial or investment advice.