Cheaper AI tokens are pushing company AI bills higher
Apollo economist Torsten Slok says falling AI token prices are driving heavier use, leaving companies with bigger model bills.
By Maya Lindqvist · Senior Technology Correspondent
3 min read
Companies are paying less for each AI token, but many are spending more on large language models as usage rises faster than prices fall. Apollo chief economist Torsten Slok says the pattern shows Jevons paradox at work in the AI boom.
The Silicon Data Token Expenditure Index says the price of a single token has fallen by more than 90% since 2023, while spending on large language models has doubled since late last year. A token is the basic unit processed by AI systems, and lower token costs can make more tasks economical to run through models.
Slok, writing in an Apollo blog post, said cheaper tokens do not lead companies to cut spending. Instead, he said, businesses use more AI agents, automate more work and generate more code, pushing total spending higher even as each unit becomes cheaper.
Jevons paradox reaches AI budgets
The economic idea Slok cited dates to William Stanley Jevons, who observed in 1865 that more efficient steam engines did not reduce coal use. As coal became cheaper to use for each task, overall demand increased.
Slok and other analysts are applying that same logic to AI. Lower per-token pricing lets companies use models more often and for more complicated work, which can keep total bills elevated.
Bain and Co. analysts reached a similar conclusion in a brief published last week. They found that token costs fell by half from December 2024 to December 2025, while the number of tokens consumed rose 450% over the same period.
Bain analysts said companies are often moving to newer AI models to gain more capability instead of keeping older systems and taking the savings. They also said AI agents handling more complex tasks tend to use more tokens per query, and teams increase their requests once they believe the systems can do more.
Companies look for limits
AI token spending has become a practical concern for companies encouraging employees to use the technology. Fortune has reported that companies including Meta and Amazon have pushed workers to adopt AI tools, while the practice of ramping up AI usage has been described as “tokenmaxxing.”
Some companies are now confronting the cost of broad deployment. Uber president and chief operating officer Andrew Macdonald recently said the company exhausted its AI budget during the first four months of the year as use of Claude Code increased, according to Fortune. Bloomberg later reported that Uber capped monthly AI spending at $1,500 per employee.
Tech executives have also pointed to compute costs as a challenge. Bryan Catanzaro, Nvidia’s vice president of applied deep learning, told Axios that compute spending for his team was far above employee costs.
The issue extends beyond software bills. Slok has also used Jevons paradox to discuss labor markets exposed to AI, citing Brookings research that generative AI could automate 86% of tasks for customer service workers. He said employment among call center workers in the Philippines has nearly doubled over the past decade, while the number of U.S. radiologists has risen 10% in 10 years despite predictions that AI would threaten the profession.
Bain analysts said corporate operating costs could shift toward a mix of 70% human headcount and 30% tokens. They warned that companies will need to measure the financial return from specific AI tools, not just set spending limits, if they want their AI budgets to hold up.
This story draws on original reporting from Fortune.