An engineer at OpenAI processed 210 billion “tokens” — enough text to fill Wikipedia 33 times — through the company’s artificial intelligence models over the last week, the most of any employee.
At Anthropic, a single user of the company’s A.I. coding system, Claude Code, racked up a bill of more than $150,000 in a month.
And at tech companies like Meta and Shopify, managers have started to factor A.I. use into performance reviews, rewarding workers who make heavy use of A.I. tools and chastening those who don’t.
This is the new reality for coders, some of the first white-collar workers to feel the effects of A.I. as it sweeps through the economy. A.I. was supposed to help tech companies boost productivity and cut costs. But it has also created an expensive new status game, known as “tokenmaxxing,” among A.I.-obsessed workers who are desperate to prove how productive they are.
At some tech companies, including Meta and OpenAI, employees compete on internal leaderboards that show how many tokens — the atomic unit of A.I. use, roughly equivalent to a word fragment — each worker consumes, two people familiar with those companies’ practices said. Generous “token budgets” are becoming a job perk for coders, like dental insurance or free lunch, and some are spending thousands of dollars a month trying to automate as much of their own work as possible.
“I probably spend more than my salary on Claude,” said Max Linder, a software engineer in Stockholm. (Mr. Linder’s employer pays for his tokens.)
Until recently, power users might have consumed thousands of tokens a day using an A.I. tool like ChatGPT, Claude or Gemini. A student writing an essay, for example, may go through 10,000 tokens — roughly equivalent to 7,500 words — including several rounds of revisions. Using millions of tokens would require hours in front of a computer, doing nothing but typing, and using billions of tokens was virtually impossible.
But the advent of so-called agentic coding tools has upped the ante. These systems can work unsupervised for hours at a time, reviewing and editing large code bases and writing entire software programs from a single prompt. Each agent can spawn subagents to handle different parts of a task, generating thousands of tokens at each step. Some A.I. systems, like the popular open-source A.I. assistant OpenClaw, are designed to run 24/7, churning through tokens while their human users sleep.
“If you have some continuously running agents, you’ll do 700 million tokens a week from a single full-time agent,” said Ege Erdil, a co-founder of Mechanize, an A.I. start-up, who estimated his own token consumption at between one billion and 10 billion a week. “It doesn’t really take that much.”
All of that adds up for the A.I. companies selling the tokens. Anthropic more than doubled its revenue projections in two months this year, largely because of the breakneck growth of its agentic coding tools. OpenAI recently said that its agentic coding tool, Codex, had tripled its weekly active users since the start of the year, and that overall Codex use, measured in tokens, had increased fivefold. Last year, Google said its A.I. models processed more than 1.3 quadrillion tokens a month.
Even for the most dedicated programmers, using billions of tokens isn’t easy. For comparison: I went through a period of heavy Claude Code use earlier this year, working on several separate coding projects for four or five hours a day, and managed to use only a few million tokens. (Rookie numbers, really.) But some coders have mastered the art of A.I. multitasking, opening multiple windows and setting dozens of agents loose on their projects at a time.
A.I. companies have encouraged these whales, giving them trophies and other rewards. And some tech executives are glad to see their employees embracing the new tools. They equate heavy A.I. use with increased productivity — if a programmer wants to operate a swarm of 10 A.I. agents, running parallel tasks in separate windows, they’re happy to foot the bill.
But I spoke to several tech workers who worried that their colleagues are gorging on billions of tokens — which can cost thousands of dollars a day — for what amount to bragging rights. Even at the A.I. labs, where workers are given unlimited use of their companies’ tools, the idea that all of this is productive seems far-fetched.
“It doesn’t seem sustainable,” said one OpenAI employee, who asked to remain anonymous because he was not authorized to discuss his colleagues’ A.I. coding addictions.
Subscribers to paid Claude and ChatGPT plans typically pay a monthly fee, which gives them a fixed number of tokens. (The number varies; some tokens are “cached,” meaning the system has stored them in memory and doesn’t need to generate them from scratch, and companies charge more for “output” tokens than “input” tokens.) Users who need more tokens can pay for them separately, or upgrade to a more expensive plan.
Shopify said in a statement that token use is just one measure of how the company measures performance. It also looks at how A.I. “improves and amplifies” work. Anthropic, Meta and OpenAI declined to comment for this column. (The New York Times has sued OpenAI and Microsoft, claiming copyright infringement of news content related to A.I. systems. The two companies have denied the suit’s claims.)
But power users have learned how to game the system by stacking multiple subscriptions or taking advantage of promotional offers. One start-up founder told me that he had discovered a loophole in an A.I. tool made by Figma, a design start-up, that allowed him to use the equivalent of $70,000 in Claude tokens through an account that costs him $20 a month. The founder, who asked to remain anonymous to avoid tipping off Figma, said he had used the loophole to build six software projects at the same time. (Figma declined to comment.)
I talked to several other tokenmaxxers about what they’re doing with all those tokens. Most were engineers or hobby programmers who were building and maintaining large, complex pieces of software using coding agents running in parallel.
They said, by and large, that A.I. coding tools were making them more productive. But some also framed their use of A.I. as a strategic move — a way to signal, to their colleagues and bosses, that they’re keeping up with the times, as the era of human coding appears to be coming to an end.
Nikunj Kothari, a venture capitalist in San Francisco, wrote in a recent Substack post about the rise of what he called “token anxiety.” He described a tech scene that has become obsessed with productivity — A.I. productivity, not human productivity — and said he had replaced Netflix with Claude Code.
“Dinner conversations used to start with ‘What are you building?’” he wrote. “That’s over. Now it’s ‘How many agents do you have running?’”
If we really are on the cusp of a white-collar job apocalypse, maybe token anxiety is rational. You don’t want to be the last programmer writing code by hand, without teams of A.I. agents working around the clock on your behalf. And employers, who are paying for all of these anxious tokens, may see it as a worthwhile expense to stay ahead of the curve.
Gergely Orosz, who writes a popular newsletter for software engineers, defended the practice of assessing workers through A.I. leaderboards, calling it “a supercheap way to learn about new and interesting ways of working.” The metrics that managers used to track programmers’ productivity before A.I. — such as how many lines of code they wrote, or how many code changes they submitted — weren’t perfect, either, he added. And for workers at the most A.I.-enthusiastic companies, Mr. Orosz said, the incentives are clear.
“Inside large tech companies, it’s becoming a career risk to not use A.I. at an accelerated pace, regardless of output quality,” he wrote.
Ah, yes, output quality. The leaderboards don’t measure that, which raises the obvious question: Are any of these tokenmaxxers producing anything good? Or are they merely spinning their wheels, churning out useless code (and wasting valuable processing power) in an attempt to look busy?
Time will tell. Maybe the A.I. addicts of today will be the 100x engineers of tomorrow. Or perhaps it’s just productivity theater — a glimmering tower of tokens, constructed by the competitive and fearful, that will topple as soon as we understand what really makes for useful work.
Either way, we’re going to need a lot more data centers.
Kevin Roose is a Times technology columnist and a host of the podcast “Hard Fork.”
The post More! More! More! Tech Workers Max Out Their A.I. Use appeared first on New York Times.
