top of page
cardiologist-doctor-surgeon-analyzing-patient-heart-testing-result-human-anatomy-interface

Your AI costs less per token. Your bill keeps going up.

Writer: Amanda
Amanda
1 day ago
6 min read

The price of using a model is falling, but each task consumes far more units than people imagine. Seven decisions to control spending before anyone notices.


Imagine the day Finance opens the first invoice for an AI agent that has been in production for weeks. Nobody knows which department requested it, which application is consuming it, or why the number is three times higher than the pilot’s. The agent works well. The problem is that nobody watched it as an expense.


It’s an increasingly likely scene, and it has a surprising cause: using AI is cheaper than ever, but operating it is not.


The token goes down. The task goes up.


AI is billed by tokens, the unit in which a model processes text. According to Gartner, a token equals about four characters. Every question sent and every answer received consumes tokens, and every token has a price.


That price is trending down. In a March 2026 press release, Gartner forecast that by 2030, running inference on a one-trillion-parameter model will cost providers more than 90% less than in 2025.


The good news comes with a warning. That drop won’t be passed along to companies in full, and agentic models, the ones that execute multi-step tasks, require 5 to 30 times more tokens per task than a standard chatbot. Because consumption is growing faster than prices are falling, Gartner expects the total cost of inference to rise.


Budgets already reflect it. A Gartner survey of 199 service and support leaders, conducted between April and May 2026, found that AI spending rose 38% while total budgets for the function grew only 2%. It’s a customer service data point, but it shows where the money is moving.


Unsurprisingly, managing AI spend is already a routine task: in the State of FinOps 2026 survey, with 1,192 professionals, 98% say they do it, up from 31% two years earlier.


The invoice doesn’t tell the whole story


The problem isn’t just how much is spent, but that it’s hard to know on what. The FinOps Foundation, a Linux Foundation program dedicated to the financial management of technology, published a guide (updated October 2026) on the cost of tokens when purchased directly from model providers. Its most revealing findings:


The invoice doesn’t say who spent. The provider shows tokens consumed by key or project, but not by department, application, or cost center. Without your own instrumentation, nobody can answer who spent what.

Buying is very easy. A developer can start consuming in minutes, with no purchasing approval or cost estimate. By the time Finance sees the first invoice, the application may already be in production.

A small oversight costs a lot. An agent stuck in a loop, an instruction that doubles in length, or a feature with ten times the expected traffic can exhaust a monthly budget in a single day. And the limits providers offer are measured in tokens per minute, not money per month.

Long conversations get more and more expensive. Agents resend their full history on every call, so the cost per call grows with the length of the session. On top of that, the response usually costs three to five times more per token than the question, and a frontier model can cost 50 to 100 times more than the smallest model from the same provider. The guide clarifies that prices change frequently and any figure should be confirmed before using it in a forecast.


And there’s a cost that doesn’t even show up on the model invoice. In production architectures with agents and information retrieval, the infrastructure surrounding the model typically accounts for 40% to 60% of total spend, according to the same guide: databases, orchestration, caches, data transfer, and monitoring.


When nobody owns the spend


Gartner describes the same pattern in software development. In a June 2026 press release, the firm predicts that by 2028 the cost of AI coding tools will exceed the average developer’s salary.


Its lead analyst, Nitish Tyagi, explains it this way: token discipline won’t emerge on its own from individual developers’ decisions, because they tend to optimize for speed and convenience before cost. Among the most common failures, Gartner cites ungoverned agent autonomy, bloated context windows, and a lack of feedback mechanisms.


The diagnosis extends to all generative AI. According to Gartner’s June 2026 report on how to optimize your costs, cited by Campus Technology, at least 50% of generative AI projects will exceed their budget through 2028 due to poor architecture decisions and a lack of operational experience. The same report warns that a production-ready system can cost orders of magnitude more than a pilot.


These are predictions, not measured data. But they all point to the same thing: the cost of AI gets discovered late.


Seven decisions before launch


These practices follow the FinOps Foundation guide and Gartner’s June 2026 recommendations. The savings percentages the foundation mentions are indicative ranges, not guarantees, and should be validated with your own data.


  1. Put a name on every expense. Every model access key should be tied to a team, an application, and a cost center, with someone accountable. It’s the minimum control and requires no additional tools.


  2. Measure by outcome, not by token. “We spent $42,000 on tokens” doesn’t help you decide. “Resolving a case costs $0.18,” as an illustrative example, does. The guide proposes measuring per query, per active user, per successful outcome, and per business transaction.


  3. Measure before budgeting. The foundation recommends observing each application for 30 to 60 days and then setting budgets with alerts, for example at 80% and 100% of the forecast.


  4. Use the smallest model that does the job. Gartner recommends sending simple, frequent tasks to small models and reserving frontier models for complex, high-value work. The foundation flags this as the lever with the greatest savings potential (60% to 90% as an indicative range). In the September 2026 State of Tokenomics survey, 86% of participants were evaluating or using a model router to do this.


  5. Send less to the model. Include only relevant information, summarize history, limit response length, reuse the fixed parts of instructions through caching, and batch-process whatever isn’t urgent. The foundation estimates savings of 20% to 60% from context management, 50% to 90% on tokens read from that cache, and around 50% with batch processing.


  6. Set caps and alerts. Gartner recommends token thresholds, escalation policies, and automatic monitoring built into the workflow. The foundation adds watching for spikes per application and not just per day, because an agent loop can consume millions of tokens in minutes.


  7. Name an owner and review regularly. In the September 2026 survey, organizations with a defined owner of AI economics were 3.7 times more likely to demonstrate value to the finance leadership. Gartner proposes periodically reviewing the workflows that consume the most tokens.


Five questions for leadership


Before approving the launch of an agent, it’s worth demanding clear answers to these five:


  1. How much does a typical run cost, and a worst-case run?

  2. Who answers for this spend, by name?

  3. Against which business outcome will we measure the cost?

  4. What cap is there per run, and what alert fires if something loops?

  5. How often do we check whether we’re still using the right model?


If any answer is “we’ll figure it out later,” that’s where the surprise on the invoice will show up.


Cost is operated too


Controlling spend isn’t a one-time project. Prices, models, and usage patterns change, and cost moves with them. That’s why it’s part of the ongoing operation, not the implementation.

At Mobiik, agent operation happens within the AI Operating Program, with 24/7 observability, FinOps, quality controls, and continuous evolution, and the cloud infrastructure is paid for directly by the client. The idea is for the cost of operating AI to be a number the business knows and governs, not a surprise.


The question that matters


It’s not how much a token costs. It’s how much it costs for your agent to resolve a case, who answers for that number, and whether you’d find out this week, and not next month, that it went up.


At Mobiik, we operate AI agents in production with continuous monitoring and spend control as part of the operation, so the cost is known before it spikes. If you want to review how much it would cost to operate your first agent and how to measure it, let’s talk.

 
 
bottom of page