AI Costs Less. So Why Are We Spending More?

Anyone using generative AI intensively eventually runs into the same issue: tokens and credits disappear much faster than expected.

This is especially true when the work becomes more complex, when several iterations are required, or when the latest and most capable models are used.

For many of us, AI has already become part of our actual work. Usage limits, credits and model costs are therefore becoming difficult to ignore.

We hear a lot about the falling cost of AI. Models are becoming more efficient. The cost of producing a given level of capability is decreasing. Competition between providers is increasing.

And yet every new model makes the previous level of usage feel outdated.

Providers naturally promote the latest model: better reasoning, better coding, better images, more realistic videos, longer context and more autonomous agents.

Even when they explain that the newest model is more expensive and that less advanced models should be used for less complex tasks, most users will still go for the latest.

Not because every task requires it, but because very few people can confidently determine which cheaper model would still be good enough for the task.

So while the cost of a specific capability may fall, the level of capability we are encouraged to consume keeps moving upward.

Image and video generation make this particularly visible. You do not pay only for the final image or video. You also consume credits for the attempts that did not work: the wrong movement, the distorted face, the inconsistent scene, or the output that looked promising but could not actually be used.

The advertised cost of one generation is therefore not necessarily the cost of producing one usable result.

At enterprise level, the issue becomes more complex. In theory, companies should use smaller and cheaper models for simple tasks, more capable models for complex work, and specialised models for code, image, video or other specific uses.

In practice, who determines that a task is “simple”?

To make that decision properly, the organisation would have to define the minimum acceptable quality, test different models on its own use cases, measure errors and human corrections, include security and compliance requirements, and repeat the assessment whenever models or prices change.

Without this capability, employees will naturally use the most powerful model available to be safe. Or the provider will make the choice through an automated routing system that the organisation does not fully understand.

This makes it increasingly important for companies to understand and track what AI is actually costing them.

When organisations evaluated major software and cloud transformations, total cost of ownership was an important part of the business case. The licence or infrastructure price was only one component. Implementation, integration, maintenance, support, training and operating costs also had to be considered.

Are companies applying the same discipline to AI?

Can procurement teams forecast expenditure when the cost depends not only on the number of users, but also on usage intensity, the models selected, the amount of context processed, the number of agentic steps and the iterations required to obtain a usable result?

Are these costs reflected in the organisation’s AI strategy? And is the value created greater than the total cost of using the tools?

A cheaper model may become expensive if employees constantly have to verify, correct or repeat its work. A frontier model may provide a slightly better result at several times the cost, without creating proportionally more business value.

The relevant question is therefore not only how much a token, request, licence or model costs. It is how much the organisation spends to obtain a result it can actually use, and whether that result creates more value than it costs.

Keeping track of AI costs should consequently become part of AI governance. Not simply to reduce spending, but to ensure that growing AI consumption is matched by measurable value.