AI models are getting cheaper.

So why could your AI bill still get bigger?

The answer is increasingly how AI gets used, not simply what each inference costs.

A traditional AI interaction might require one request and one response. An agentic workflow can involve reasoning, retrieval, tool calls, validation, retries, and multiple model interactions before the task is complete.

Recent industry research projects inference costs per agentic workflow to increase more than fivefold through 2028, even as model-level economics continue to improve.

Stop Measuring AI Cost Per Token

Token and inference pricing still matter. But they don’t answer the question that matters most to a CTO:

What does it actually cost to successfully complete the work?

A $0.01 inference isn’t necessarily cheaper than a $0.03 inference if the first requires twelve steps and the second completes the task in four.

The unit price is different. The workflow economics may not be.

Measure the Cost of the Whole Workflow

This is where enterprises need a different lens:

Inference + Tool/API Calls + Data/Retrieval + Infrastructure + Retries + Human Intervention

↓

Cost per Successful Workflow

↓

Business Outcome → ROI

The objective isn’t to eliminate every cost. It’s to understand where consumption accumulates and whether the resulting business outcome justifies it.

The Cheap Model Trap

Lower inference prices can make more complex AI workflows economically viable.

That can encourage more reasoning, more automation, and more steps.

The result is an important paradox:

Lower cost per step can lead to higher total consumption.

That doesn’t mean cheaper models are bad—or that agentic AI is inherently expensive.

It means model price alone is an incomplete measure of AI economics.

Scale What Makes Economic Sense

Before scaling an agentic workflow, CTOs and CIOs should ask:

  • What does one successful workflow cost?
  • How many inference and tool calls does it require?
  • Where does most of the cost accumulate?
  • How often does it retry or require human intervention?
  • What measurable value does each successful workflow create?

The goal isn’t to find the cheapest model.

It’s to build the most economically efficient way to complete the work.

Because in agentic AI, the real unit of value isn’t the inference.

It’s the outcome.