The latest challenge in artificial-intelligence adoption is "tokenomics" and the management of AI expenses that can add up quickly if customers aren't careful.
Tokens are the basic unit of AI input and output. For example, this sentence has 17 tokens when counted by OpenAI's models. Typically, 100 tokens represents about 75 words. You can try AI token counting for yourself on OpenAI's site.
The token has become the currency of the AI age. The problem is tokens have a different value depending on how they're used. Customers are confused because they can't set budgets, and Wall Street is puzzled because it can't easily track pricing the way they can in retail stores or even the cloud.
Anthropic complicates matters because it counts tokens differently than the rest of the AI world. The firm recently changed how some of its products tokenize text, meaning the same words use 30% to 40% more tokens than they did before, even while its older models still use the more standard method of counting.
Consumers have become accustomed to all-you-can-use plans costing $20 from both ChatGPT and Anthropic's Claude, but enterprises pay by the token. For the best ChatGPT model, the cost ranges from 50 cents to $30 per one million tokens.
On the surface, that doesn't sound like much, but two things have changed the math and turned AI into an expensive habit.
The first was the rise of "reasoning" models in 2024, which gave chatbots a more introspective approach to answers. Unlike the quick responses of early ChatGPT, reasoning models talk to themselves and work through a problem step-by-step. In preparation for this article, I asked ChatGPT to "tell me about all the stories of companies overdoing it with token spending." The model had a 32-second conversation with itself in which it performed multiple web searches and read dozens of articles before coming up with a thousand-token response. The internal conversation used up much more than that.
The more recent -- and bigger issue -- is the rise of AI agents, which can accomplish a complex series of tasks from a simple conversational command. Unlike humans, restrained by human limits, agents chew through tokens. Every prompt generates a flood of conversations -- requiring ever more tokens, all happening without human intervention.
An April study from researchers at Alphabet's Google, Microsoft, and top universities found that coding agents generate over a thousand times more tokens than people do for the same tasks. Today, agents are most commonly used for software coding, and the researchers found that token costs are highly variable task-to-task and model-to-model. Both people and AI models are bad at estimating how many tokens a job will consume, the study found. Think of it like having a $20 bill in your wallet and not knowing how many hot dogs it can buy.
More tokens don't necessarily lead to better results, and the jobs that fail have to be redone, at the cost of yet more tokens.
If trends continue, agents will soon outnumber people on enterprise networks, before eventually dwarfing human users. It's the anticipated growth of agents, and the manner in which they gobble tokens, that explains why hundreds of billions of dollars are still pouring in to build new AI data centers, with more planned for next year.
Rapid advancements in coding agents have pushed CEOs to embrace their usage, perhaps before anyone understood their cost. Earlier this year, companies created dashboards to track employee usage. This had the predictable result of huge bills followed by a hasty retreat to a more frugal approach.
The budget worries come as companies continue to debate the utility of AI itself. In a May podcast interview, Uber Technologies President Andrew Macdonald captured the new caution. "How many projects that were on the cutting room floor got moved above the line because of the productivity gains?'" he said. "That link is not there yet."
Meanwhile, Wall Street is dealing with its own crisis of confidence when it comes to AI monetization. Analysts would like to use tokens as a proxy for the health of the AI trade -- but all of the issues above are clouding the picture.
As costs rise for the best U.S. frontier models, companies are finding other ways to control their AI budgets, which means going down the price curve, including using ultracheap models from China. Eventually, a more competitive market could force Anthropic and OpenAI to cut prices on their top models, but we're not there yet.
For now, the best way to track the balance of supply and demand for AI is via the cloud. Investors should pay more attention to pricing at Amazon Web Services, Microsoft Azure, and Google Cloud. Just this week, AWS raised prices for Nvidia GPU rentals for the second time this year.
Ultimately, for AI to be viable it has to bring increased productivity, which is essentially profit divided by hours of human work. The labor costs are well understood, but the profit part of the equation requires a better understanding of token costs.
Until then, companies will do everything to keep token spending in check. And that could be bad news for beneficiaries of the AI economy, from chip makers to energy firms to model providers like OpenAI and Anthropic.
As Stanford economist Erik Brynjolfsson told me last year, "Technology by itself rarely delivers productivity just when you plug it in. What almost always has to happen is that you have to rethink your business processes," he said. "It could take years or decades to really work through all the possibilities."
Write to Adam Levine at adam.levine@barrons.com
This content was created by Barron's, which is operated by Dow Jones & Co. Barron's is published independently from Dow Jones Newswires and The Wall Street Journal.
(END) Dow Jones Newswires
July 02, 2026 03:00 ET (07:00 GMT)
Copyright (c) 2026 Dow Jones & Company, Inc.
Comments