One of the most persistent assumptions about generative AI is that it will inevitably become cheap.

Models will become more efficient. Hardware will improve. Competition will increase. Smaller models will handle tasks that once required frontier models. The cost of a million tokens will continue to fall.

All of that is probably true.

But I increasingly think it leads us towards the wrong question.

The interesting question is no longer:

When will AI models become cheap?

They already are becoming cheaper on many useful measures.

The more important question is:

When will the AI work we actually want done become cheap?

Those are not the same thing.

The price of intelligence is falling

At the model level, the economics are moving quickly.

Providers continue to improve the amount of intelligence available for a given dollar. New generations of models routinely deliver capabilities that previously required more expensive models, and competition between OpenAI, Anthropic, Google and others continues to put pressure on inference costs.

The result is that the cost of an individual model call can become surprisingly small.

For many traditional AI applications, this matters enormously.

Classification, summarisation, extraction, routing and straightforward question answering can often be performed for fractions of a cent or a few cents.

This encourages a simple mental model:

But the second conclusion does not necessarily follow from the first.

We are changing what we ask AI to do

The problem is that the unit of AI consumption is changing.

A few years ago, a typical interaction might have been:

Summarise this document.

The model receives some tokens and produces some tokens.

The economics are easy to understand.

Increasingly, however, we ask AI systems to do something more like:

Research this subject, inspect these documents, find supporting evidence, analyse the alternatives, create a recommendation, generate a PowerPoint presentation and check the result.

That isn't really a single model call.

It is a computational process.

The agent may reason about the problem, search for information, retrieve documents, call tools, execute code, inspect intermediate results, revise its approach and potentially invoke the model many times before returning anything to the user.

The economics therefore start to look very different.

The model may be cheaper while the amount of model usage, tooling and computation devoted to the task increases substantially.

A PowerPoint can already cost dollars, not cents

I recently experimented with using an AI agent to create a PowerPoint presentation.

The task was not simply to generate some slide text. The agent researched the subject, reasoned about the structure, generated content, assembled the presentation and iterated on the result.

The usage I observed worked out at roughly US$8 for the task.

That should not be interpreted as a fixed price for an AI-generated PowerPoint. Agentic workloads vary substantially depending on the model, the amount of context, the number of iterations, tool usage and the complexity of the task.

But that is precisely the point.

The economically relevant unit is increasingly not:

How much did one model response cost?

It is:

How much computation did the system consume to complete the task?

An $8 task is not necessarily expensive.

If it replaces several hours of human work, it may be extraordinarily cheap.

But it is very different from the assumption that widespread enterprise AI will consist primarily of negligible fractions-of-a-cent transactions.

The industry is starting to price the task, not just the token

We can already see this shift appearing commercially.

Traditional model APIs largely expose token-based economics. Agentic products increasingly have to account for much more than tokens.

Microsoft's Copilot Cowork, for example, uses usage-based pricing where the cost of a task reflects model use, organisational context retrieval, tool calls and runtime.

That is an important development because it makes explicit what was previously hidden inside an application architecture.

The commercial metric is beginning to move closer to the actual economic activity being performed.

Three different AI cost curves

This is why discussions about whether AI is becoming cheaper can be confusing.

There are actually at least three different cost curves.

All three things can happen simultaneously.

The cost of intelligence can collapse while organisations spend more money on AI than ever before.

The Jevons paradox may apply to intelligence

There is a useful economic analogy here, although it should not be treated as a prediction.

The Jevons paradox describes situations where improvements in the efficiency of using a resource reduce its effective cost and encourage enough additional consumption that total usage increases rather than decreases.

AI may exhibit similar behaviour.

As intelligence becomes cheaper, we are unlikely to keep doing exactly the same amount of reasoning at lower cost.

We will probably use more reasoning.

We will analyse things that were previously not worth analysing.

We will generate multiple alternatives instead of one.

We will continuously monitor systems that humans previously checked occasionally.

We will ask agents to review thousands of documents instead of sampling twenty.

We will run simulations, evaluations and verification processes that would previously have been economically impractical.

And agents themselves may consume additional intelligence deciding what intelligence to use.

Whether AI ultimately demonstrates a textbook Jevons paradox remains to be seen.

But the underlying mechanism is already plausible: falling unit costs can unlock enough new demand that total expenditure does not fall proportionally.

Agents change the economics again

Agentic systems amplify this effect because they introduce something that ordinary software rarely possessed:

the ability to decide dynamically how much computation to spend solving a problem.

A traditional application executes a relatively predictable sequence of operations.

An agent may decide that it needs another search, another document, another model call, another tool invocation or another attempt.

More capable agents can therefore consume more resources precisely because they are capable of doing more work.

Features such as extended reasoning, larger context windows, computer use, code execution, web research and multi-agent delegation expand what a system can accomplish, but they also expand the number of ways it can consume resources.

That makes agentic AI economics fundamentally different from simply calculating a token price.

Multi-agent systems make this even more interesting

Now consider multi-agent architectures.

One agent receives a complex task and delegates independent parts to several other agents.

Each subagent may have its own context, reasoning process, searches and tool calls.

The parent agent then evaluates and synthesises their work.

This can be enormously valuable when parallelism or independent reasoning improves the result.

But it also means a seemingly simple user request can trigger a significant amount of invisible computation.

This reinforces an important architectural principle:

Enterprise AI will need its own FinOps discipline

Cloud computing went through a similar learning process.

The cloud initially appeared to make infrastructure dramatically cheaper and easier to consume.

It did.

But it also made infrastructure dramatically easier to provision.

Organisations subsequently discovered that consumption without visibility or governance could create surprisingly large bills.

That led to the emergence of FinOps practices focused on understanding usage, allocating costs and continuously optimising cloud expenditure.

AI is likely to require a related discipline, although the optimisation problem is different.

The objective should not simply be to minimise token consumption.

It should be to understand the relationship between:

The cheapest model is not necessarily the cheapest solution.

A more expensive model that solves a task correctly on the first attempt may cost less overall than a cheaper model that requires retries, additional validation or human correction.

Likewise, spending $10 on an agent that eliminates $300 of manual effort is economically sensible. Spending $1 on an agent performing a task worth 20 cents is not.

This makes AI cost optimisation inseparable from measuring business value.

And that measurement is going to be difficult

This may be one of the less discussed challenges.

Infrastructure costs are comparatively easy to measure.

Business value is not.

An agent might save an employee twenty minutes, produce a better decision, avoid an error, accelerate a customer response or discover something that would otherwise have been missed.

Those benefits are real, but they are not always easily converted into a dollar value.

At the same time, the cost side may include more than the visible AI consumption.

Enterprises also need to consider integration, evaluation, monitoring, security, governance, human review and the operational cost of correcting failures.

A mature view of AI economics therefore needs to consider the whole system, not merely the API invoice.

The future probably isn't cheap AI

I suspect the phrase "AI will become cheap" will eventually look slightly misleading.

A better description might be:

Intelligence will become cheap enough that we use vastly more of it.

Individual units of reasoning will continue to fall in price.

Smaller models will perform tasks that currently require frontier models. Hardware and inference systems will improve. Competition will continue to push prices down.

But at the same time, the amount of useful work we ask AI systems to perform is likely to expand enormously.

The future enterprise may have agents continuously researching, analysing, monitoring, generating, validating and acting across thousands of business processes.

Some individual tasks may cost almost nothing.

Others may consume dollars, tens of dollars or considerably more because the economic value of completing them justifies that expenditure.

So perhaps the question isn't:

When will AI get cheap?

Perhaps it is:

How much intelligence will become economically worthwhile once intelligence is cheap?

The answer to that question could be enormous.

And if it is, falling AI prices may not lead to smaller AI bills at all.

They may be what makes much larger ones worthwhile.