san francisco: OpenAI engineers processed 210 billion “tokens” (enough text to fill Wikipedia 33 times) through the company’s artificial intelligence models in one week this month. This is the highest number among all employees.
At Anthropic, one user of Claude Code, the company’s AI coding system, collected more than US$150,000 (RM598,950) in bills in one month.
And at technology companies like Meta and Shopify, managers have begun to factor the use of AI into performance reviews, rewarding employees who use AI tools frequently and disciplining those who don’t.
This is a new reality for programmers, who are some of the first white-collar workers to feel the impact of AI as it permeates the economy. AI was supposed to help technology companies increase productivity and reduce costs. But it has also created an expensive new status game known as “tokenmaxxing” among AI-obsessed workers desperate to prove how productive they are.
At some tech companies, such as Meta and OpenAI, employees compete on internal leaderboards that show how many tokens (atomic units in AI usage, roughly equivalent to word fragments) each employee consumes, said two people familiar with the companies’ practices. Generous “token budgets” such as dental insurance and free lunches are becoming job perks for programmers, with some spending thousands of dollars a month to automate as much of their work as possible.
“I probably spend more on Claude than my salary,” said Max Linder, a software engineer in Stockholm. (Linder’s employer pays for his tokens.)
Until recently, power users could have spent thousands of tokens a day using AI tools like ChatGPT, Claude, and Gemini. For example, a student writing an essay may go through 10,000 tokens (equivalent to approximately 7,500 words), including several rounds of revision. Spending millions of tokens required hours of doing anything other than typing in front of a computer, making spending billions of tokens virtually impossible.
But the bar has been raised with the advent of so-called agent coding tools. These systems can run unsupervised for hours at a time, review and edit large code bases, and create entire software programs from a single prompt. Each agent can spawn subagents to handle different parts of the task, generating thousands of tokens at each step. Some AI systems, such as the popular open-source AI assistant OpenClaw, are designed to run 24/7, generating large amounts of tokens while human users are asleep.
“If you have agents running continuously, you’re processing 700 million tokens a week from one full-time agent,” said Ege Erdil, co-founder of AI startup Mechanize, who estimates his token consumption at 1 billion to 10 billion a week. “It actually doesn’t take that long.”
All of this is a plus for AI companies selling tokens. Anthropic more than doubled its revenue estimates in two months this year. This is primarily due to the tremendous growth in agent coding tools. OpenAI recently announced that the number of weekly active users of its agent coding tool, Codex, has tripled since the beginning of the year, and that overall Codex usage, measured in tokens, has increased five times. Last year, Google announced that its AI models processed more than 13,000 trillion tokens per month.
Working with billions of tokens is not easy, even for the most dedicated programmers. For comparison: I used Claude Code a lot earlier this year, and there was a period where I was working on several separate coding projects for 4-5 hours a day, and I managed to spend only a few million tokens. (Actually, that’s the number of newbies.) But some programmers have mastered the art of AI multitasking, opening multiple windows and unleashing dozens of agents on a project at once.
AI companies have encouraged these whales, giving them trophies and other rewards. And some technology executives are happy to see their employees taking advantage of new tools. They equate heavy use of AI with increased productivity. If a programmer wants to work with a swarm of 10 AI agents performing parallel tasks in separate windows, we’re willing to pay for it.
But I spoke with several tech workers who were worried that their co-workers were gobbling up billions of tokens, sometimes costing thousands of dollars a day, for what amounted to bragging rights. The idea that all this could be productive, even in an AI lab where employees have unlimited use of company tools, seems far-fetched.
“I don’t think that’s sustainable,” said one OpenAI employee, who requested anonymity because he was not authorized to discuss his colleague’s AI coding addiction.
Subscribers to paid Claude and ChatGPT plans typically pay a monthly fee and are granted a fixed number of tokens. (The number varies; some tokens are “cached,” meaning the system stores them in memory and doesn’t need to generate them from scratch. Companies also charge more for “output” tokens than for “input” tokens.) Users who want more tokens can pay separately or upgrade to a more expensive plan.
Shopify said in a statement that token usage is just one measure of how the company measures performance. We also look at how AI can “improve and scale” jobs. Anthropic, Meta, and OpenAI declined to comment for this column. (The New York Times sued OpenAI and Microsoft, alleging copyright infringement of news content related to its AI systems. Both companies denied the lawsuit’s allegations.)
I talked to some other tokenmaxxers about what they’re doing with all these tokens. Most were engineers or hobby programmers who built and maintained large, complex software using coding agents running in parallel.
They said that, in general, AI coding tools have increased their productivity. But some see the use of AI as a strategic tool, a way to let colleagues and bosses know they’re keeping up with the times as the era of human coding appears to be drawing to a close.
If we truly are on the brink of a white-collar employment collapse, perhaps token anxiety is reasonable. You don’t want to be the last programmer writing code by hand without a team of AI agents working around the clock on your behalf. And employers paying these uncertain costs may see it as a worthwhile expense to stay ahead of the curve.
Gergely Orosz, who writes a popular newsletter for software engineers, defends the practice of evaluating workers through AI leaderboards, calling it “a very inexpensive way to learn about new and interesting ways of working.” Metrics that managers used to track programmer productivity before AI, such as the number of lines of code a programmer wrote or the number of code changes submitted, were also not perfect, he added. And for employees at companies most passionate about AI, the incentives are clear, Orosz said.
“Within large technology companies, it is becoming a risk for carriers not to use AI at an accelerating rate, regardless of the quality of the output,” he wrote.
Ah, yes, the output quality. Leaderboards don’t measure that, which begs the obvious question: Are any of these tokenmaxxers producing anything good, or are they just sitting idle and producing a lot of useless code in an attempt to look busy (and wasting valuable processing power)?
Time will tell. Perhaps today’s AI junkie will be 100 times more of an engineer than tomorrow. Or maybe it’s just productivity theater, a glittering tower of tokens built by competitive and fearful people that falls over as soon as they realize what actually useful work is.
In any case, more data centers will be needed. – ©2026 The New York Times Company
This article was originally published in The New York Times.
