The Current Token Economics Will Not Last

There is something peculiar about building a business whose gross margin depends on how long a machine spends talking to itself. Somewhere, a founder is celebrating the automation of an entire department. Somewhere else, an agent has spent forty-three steps reconsidering the name of a folder. Both events appear on the same invoice.

The appeal is obvious. Software that once required substantial human effort can now be assembled, interrogated and revised through a conversation. The bill arrives in tokens, small enough individually to make almost any experiment seem reasonable. Then somebody connects the experiment to a loop.

The current economics will not last. That is a prediction about how we buy and package this capability, not a claim that every model provider is secretly selling each request at a loss. Public prices do not reveal enough to establish that. The more interesting problem is that the unit being sold and the result being bought are drifting apart.

Cheap compared with what?

The strongest objection comes first: inference has become dramatically cheaper. Stanford’s 2025 AI Index reported that the price of querying a model matching GPT-3.5’s performance on the MMLU benchmark fell from $20 per million tokens in November 2022 to $0.07 by October 2024.

That is a reduction of more than 280 times. It compares a particular benchmark capability across models, not the cost of every task, and it measures prices rather than providers’ complete production costs. Nevertheless, it is substantial progress. An argument that depends on AI remaining inherently expensive has some explaining to do.

But cheaper units do not guarantee a smaller bill. If a task uses ten times as many units after the unit price halves, expenditure rises fivefold. No mysterious economic paradox is required. You have simply bought more of the thing that became cheaper.

A short answer and an autonomous investigation are different products. The latter may involve searching, retrieving documents, executing code and trying again. A falling price per token can coexist quite comfortably with an expanding appetite for tokens. Cheap electricity would not make it economical to leave the immersion heater running until it developed a philosophy.

The meter measures activity

Tokens are a useful billing unit. They provide a measurable way to charge for model usage. But the customer rarely wants tokens themselves. They want a correct answer, a working change or a completed piece of administration.

This distinction matters when a system can spend more without achieving more. An agent that fails after thirty attempts has been extremely active. So has a printer consuming an entire cartridge to explain that it cannot connect to the network.

Researchers have identified this problem in evaluation. In AI Agents That Matter, Sayash Kapoor and colleagues argue that accuracy alone is an inadequate measure of agent performance. Their experiments show the importance of considering cost alongside results, including cases where simpler approaches compete with more elaborate systems.

That does not establish that every complicated agent is wasteful. It establishes that sophistication needs to earn its invoice. A diagram containing six agents, three critics and a supervisor is not evidence of efficiency. It may be an unusually expensive way to ask the same question several times.

For a developer, the relevant calculation is the cost of producing an acceptable result, including failed attempts and review. The headline token rate is one ingredient. Treating it as the entire calculation is like choosing a builder by the price of the nails.

A token already has several prices

Even the tariff is more complicated than the language suggests. Anthropic’s published pricing distinguishes input from output and cache reads from cache writes. It also lists separate charges for services such as web search and code execution.

These are provider terms, not independent evidence about profitability. They do demonstrate that “the token price” is already an incomplete description of the bill. Reusing material, generating new material and operating tools are different activities with different charges.

That creates opportunities for good engineering. Repeated context can sometimes be cached. Work can be routed to a less expensive model when its performance is sufficient. A deterministic operation need not become a conversation merely because the conversation is available.

It also creates a less glamorous responsibility: somebody must understand the bill. The product may promise to remove administrative work while quietly creating a new specialist in why Tuesday cost more than Monday.

The subscription meets the enthusiast

A consumer subscription and a metered API solve different commercial problems. A subscription sells predictable access within defined limits. An API passes usage costs through more directly. Trouble begins when a product promises something resembling the former while buying its essential input through the latter.

Consider a hypothetical service charging £20 a month. One customer asks it to summarise two documents. Another uses it continuously to investigate an archive. Equal revenue does not imply equal cost. The enthusiastic customer may also be the one most delighted with the product, which makes the resulting conversation particularly awkward.

There are reasonable responses: usage allowances, higher tiers, separate charges or limits on expensive operations. None is evidence of fraud. They are ways of deciding who bears the cost of unusually heavy demand.

But they undermine the assumption that a modest fixed payment can indefinitely purchase an expanding amount of autonomous work. A human pauses to eat lunch. A loop requires an explicit instruction.

The infrastructure is not imaginary

The physical investment behind these services is substantial. In February 2026, Associated Press reported that Alphabet expected capital expenditure of $175 billion to $185 billion for the year, following $91 billion in 2025.

Those figures cover broad infrastructure investment. They cannot tell us the profit or loss on a particular model request, and dividing them by an arbitrary number of tokens would produce theatre rather than accounting. They do remind us that falling retail prices sit above real equipment, financing and operating requirements.

Efficiency can improve. Demand can grow. Some businesses may earn excellent returns. The mistake is treating any one of those possibilities as proof that today’s product packaging has reached its final form.

What survives the adjustment

My expectation is that ordinary tasks will keep getting cheaper while reliable, demanding work remains differentiated by capability, speed and service. Some products will charge by usage. Others will absorb that usage into subscriptions, workflow fees or prices attached to completed tasks.

Token billing may survive quite happily underneath. Businesses still pay for computing resources their customers never think about. The more fragile proposition is that customers should organise their understanding of value around the machine’s internal activity.

Charging for outcomes has difficulties too. Somebody must define success, handle disputes and account for work whose value emerges later. “Completed” becomes a lively word when a customer discovers that the automation has confidently completed the wrong thing.

There is no tariff that abolishes the need for a sound business. The durable products will have to make useful work cheaper or better, measure that honestly and leave enough margin to continue doing it.

The founder’s dashboard may still celebrate millions of tokens processed. The customer will have a smaller question: did the thing get done? Eventually, that question will reach the person paying for the GPUs. It is unlikely to be impressed by how much the machine had to say.