SAN FRANCISCO — Earlier this year, the message from technology companies to their employees was clear: use as much artificial intelligence as possible at work.
Employees called it “tokenmaxxing”with “token” referring to a unit of AI usage, roughly equivalent to a word fragment. Meta and Amazon employees even competed in rankings that tracked the use of tokens.
Then the bills came in from companies like Anthropic and OpenAI that provide AI tools — and they weren’t cheap. Now, the era of tokenmaxxing appears to be over.
Meta told employees last week that it would soon limit its use of AI, after observing a “exponential increase” in costs. In May, Uber stated that it had consumed the entire projected budget for AI spending for the year in just four months and began imposing some monthly limits on AI programming tools. Walmart has also set limits on different AI tools. And Amazon and Meta have taken down tokenmaxxing rankings.
In other words, “tokenminting”short for “token minimizing” (“token minimization”), is now trending.
This turnaround, in just a few months, shows how the use of AI is still in flux, as people and companies try to figure out how best to use these tools.
Continues after advertising
“The biggest problem is that everything is changing so fast that people and companies don’t know what to do”said Rob May, CEO of Neurometric, a startup that helps companies use AI better, and author of “The Tokenminning Manifesto”.
“CEOs who didn’t know how to measure their employees’ familiarity with AI thought, ‘Well, who’s using the most tokens?’”he said, adding that this philosophy ended up promoting volume over efficiency.
OpenAI and Anthropic offer subscriptions that cost $10 to $200 per month for use of its AI models; When subscribers reach their usage limit, access stops. But most of the revenue comes from offering tools to companies like Meta, Shopify and Amazon, which pay not just subscription fees but also for the tokens used by their tens of thousands of employees. Thus, the more tokens are used, the more expensive AI becomes.
Continues after advertising
A simple task, like asking AI to summarize the transcript of a company meeting, can consume a few hundred tokens. More complex requests, such as writing code to create a new product or feature, can use tens of thousands.
The costs of using AI models have skyrocketed as they have become more powerful and consumed more tokens. Anthropic’s latest model, Fableis twice as expensive than the previous model, Opus. While there are cheaper models, many employees have gotten into the habit of using the more powerful models for everything, May said.
The ways in which AI is used have also changed. Instead of just chatting with chatbots, engineers started using AI “agents”capable of working on complex tasks for hours at a time. As a result, engineers can consume tens of thousands of dollars in tokens per month.
Continues after advertising
Many companies said they were trying to be more strategic with their AI spending after not seeing clear returns on that investment.
“If you can’t draw a direct line between that and the amount of useful functionality you’re delivering, that tradeoff becomes harder to justify”said Andrew Macdonald, Uber’s chief operating officer, in a recent podcast interview. “That connection doesn’t exist yet.”
This doesn’t mean that companies will stop spending heavily on AI. Meta told employees he was on track to spend billion with the use of AI this year, but I wanted “find areas where we can spend less achieving similar or better business results”. Marc Benioff, CEO of enterprise software company Salesforce, said his company planned to spend hundreds of millions with AI this year, but now monitors “agentic work units” instead of tokens. The new metric aims to measure production, not just use.
Continues after advertising
Meta and Walmart’s limits on the use of AI by employees had previously been reported by The Information and by Bloomberg.
It is still unclear how the tokenminting could affect Anthropic and OpenAI’s financial results. At the height of the tokenmaxxing era this year, AI companies reported record revenues driven by the use of programming tools. Last week, Meta told its engineers to use its internal programming assistant, the MetaCoderather than third-party tools whenever possible.
Meta declined to comment, Anthropic did not provide comment, and OpenAI did not respond to a request for comment. (The New York Times has sued OpenAI and Microsoft, alleging copyright infringement of journalistic content related to AI systems. The companies have denied the allegations.)
The clearest path for companies going forward, May said, is to use cutting-edge AI only in complex tasks that really require it, substituting cheaper models in other cases.
Companies can save up to 90% by opting for less advanced AI models, said Andy Markus, chief AI officer at AT&T. According to him, his engineers use the most powerful models for some tasks and the least powerful for most other activities.
“There is a back and forth”he said. “What we realized is that for most use cases, the newer, more advanced frontier model is not necessary.”
c.2026 The New York Times Company
