When Token becomes the performance metric, employees are herded into a new racetrack
A friend of mine working in tech at a large company told me two stories over dinner last week.
The first one was about his own company: during this half-year of layoffs, the company issued every remaining employee a free token quota, claiming it was a gesture of reassurance, but it was essentially a disguised mandate—no one could afford not to use AI. Behind this "free" benefit, however, lies a hidden ledger: every employee's token usage volume and frequency are tracked, recorded, and ranked.
The second story was something he heard: several other companies do not cover this cost at all. Employees who want to keep up with AI have to pay for subscriptions out of their own pockets. More expensive plans can cost up to 2,000 yuan a month—exactly the price of a month's rent in a city like Shanghai. Even cheaper plans run 200 to 300 yuan a month, enough for a decent group dinner. A group of tech professionals around him have already started jokingly calling themselves "Chat Engineers": their daily work consists of typing into dialog boxes, staring at progress bars, and waiting for AI to finish the rest. What makes them even more anxious is a phrase circulating around the office: by 2027, the programming job as we know it might no longer exist.
This approach is not unique to their company. Tech giants on the other side of the Pacific have taken it even further. Janelle Gale, Meta's Chief People Officer, previously told employees that "AI-driven impact" would be one of the company's key performance evaluation priorities for 2026. Andrew Bosworth, Meta's Chief Technology Officer, mentioned at a tech summit that one of his engineers spent as much on AI tokens in a single year as his own annual salary, which yielded 5 to 10 times the output. His exact words were that it was almost like getting free money: "Keep going, no upper limits."
A few weeks later, an internal Meta leaderboard was created to rank employees by their token usage: an employee-developed tool called "Claudeonomics" covered more than 85,000 employees across the company, listing the top 250 users by total token consumption, and awarding top performers playful titles such as "Token Legend" and "Cache Wizard." According to data from The Information cited by Fortune, over a 30-day period, the total recorded consumption exceeded 60 trillion tokens, with the top-ranked employee alone using around 281 billion tokens. Using the then-rate of $5 per million tokens for Claude Opus 4.6 as a rough estimate, this volume would theoretically be worth more than $1.4 million. The actual amount Meta paid has not been disclosed. The leaderboard was shut down two days after it was exposed, but a detail before its closure is more thought-provoking than the data itself: Bosworth, who advocated for "no upper limits," and Mark Zuckerberg both failed to make it into the top 250 of this leaderboard built by their own employees.
This reveals a deeper truth: companies cannot clearly quantify how much value AI has actually created, so usage volume itself has become a substitute for value. Instead of being replaced by AI, employees have been herded into a new track—one with no finish line, no referees, and even those who advocate "no upper limits" are not leading the pack.
Unable to calculate exactly what value AI creates, companies choose to count how much it is used
The reason enterprises have started tracking token usage ultimately stems from an unresolved management dilemma. When leadership mandates "full embrace of AI," the company itself hits a wall first: it cannot answer a more fundamental question—how much real value do these tools actually generate? Whether employees know how to use them becomes a secondary concern.
A feature launched a few days earlier could be credited to AI, or it could simply be due to simplified requirements. An engineer submitting more code does not necessarily mean fewer bugs. A project that originally required three people to complete now being handled by one person cannot rule out the influence of business changes or individual capabilities. The contribution of AI is mixed in with trivial tasks like information retrieval, discussions, coding, and decision-making, making it almost impossible to separate into a clear, clean bill.
Tokens, on the other hand, are honest and easy to track: every call increments the number, and every hour an agent runs pushes the usage curve upward. They can tell management whether AI is being used, but they cannot tell whether work has actually improved as a result. Managers ultimately chose this metric not because it is more accurate, but because it is more convenient. A judgment that should have been made jointly by supervisors, reviewers, and post-project retrospectives has been simplified into a single data export. There is a well-known concept in economics called Goodhart's Law: when a metric becomes a target for important decisions, people will find ways to game it, and eventually the very thing the metric was meant to measure gets distorted by those gaming the system.
Employees have seen through this game,
but they calculated that cooperating is more cost-effective
In May, an internal Amazon service called "KiroRank" scored employees based on their activity on the Kiro development platform. According to the Financial Times, some employees quickly found loopholes: they could let autonomous agents run unnecessary tasks to inflate their rankings. The longer the agents ran, the more impressive their token performance became—and the larger the bill Amazon had to pay. Amazon later shut down the service, with a senior vice president reminding employees not to use AI for the sake of using AI.
The problem is that almost no employee dares to bet on opting out of this performance. The math is simple: this evaluation is not being set by a single company alone—more and more employers are asking the same question: Are you keeping up with AI? Switching jobs will not let you exit; it just lands you on an identical leaderboard where you have to re-rank yourself. The cost of playing along is spending a little extra money—either your own or the company's. The cost of refusing, however, could be being labeled as falling behind, which would land you in the group targeted for layoffs. Comparing these two outcomes, the answer is obvious. Ultimately, employees' professional judgment has not changed; what has changed is that the rules leave them no exit option.
The more important question to ask is,
why no company dares to require employees to "pay for their own computers,"
yet some dare to demand "self-funded subscriptions"
Companies that provide tokens are recording every trace of their employees' AI usage on their ledgers. Companies that do not provide tokens are shifting the cost of entry to this track directly to their employees. These two approaches seem different, but they are essentially enforcing the same thing: prove that you are still on the track.
The truly crucial point lies in the "2,000-yuan self-payment" my friend mentioned. Requiring employees to pay for their own computers or office software would be seen as absurd in any company. Employees would rightly assume that the company covers these costs, as these are production resources that should be borne by the business. Yet when it comes to tokens—another essential resource for getting work done—some companies have nonchalantly shifted the cost to their employees' own wallets. At least among my friend's colleagues, many have accepted this, and do not even think it is unfair.
The difference lies in the appearance of this expense. A computer is a one-time purchase, clearly accounted for as a fixed asset of the company. AI tools, however, are often billed via monthly subscriptions, with overage charges calculated by the number of calls or tokens. This payment model looks nearly identical to how an individual pays for a streaming service subscription or an annual gym membership. The question of "who should pay for this"—which should have a clear answer—has been quietly blurred by this superficial resemblance. Companies do not need to explicitly point this out, and employees themselves struggle to justify why this seemingly "personal expense" should be reimbursed by the company. This cost-shifting works entirely because the expense looks enough like "personal consumption" and not enough like "company overhead."
When layoffs loom,
gamified leaderboards stop being just games
If a company is still expanding, the token leaderboard might be treated as an internal game to encourage experimentation. But once rumors of layoffs start circulating, every ranking immediately takes on a new meaning. My friend has already experienced this firsthand.
In May this year, Meta announced layoffs of nearly 8,000 employees, accounting for roughly 10% of its global workforce. In July, 26 employees in the U.S. filed a lawsuit against Meta, alleging that the company relied on AI tools to measure productivity and token usage, and included "AI adoption" in performance evaluations during the layoff process. According to Reuters, the plaintiffs claimed that the system did not pause tracking when employees took legally protected sick leave, maternity leave, or family care leave, which dragged down their activity scores and potentially made them more vulnerable to layoffs. Meta denied any wrongdoing, insisting that layoff decisions were made by humans; a U.S. federal judge rejected the employees' request for an emergency injunction to halt the layoffs, and the substantive disputes in the case are still pending arbitration.
The allegations have not been legally confirmed, and the employee-built leaderboard inside Meta cannot be directly equated to a layoff tool. But this lawsuit has laid bare a reality: tokens, online activity, and AI usage traces, once entered into a company's data system, can instantly turn from harmless records into criteria that decide who stays and who goes. Whether companies are willing to admit it or not does not change this: as long as "AI usage" is visible, and the company is conducting layoffs, this track is already active. Everyone has to keep running—not because there is guaranteed reward for finishing strong, but because stopping could mark you as someone who has fallen behind.
My friend and his peers, who jokingly call themselves "Chat Engineers," are all worried that the programming profession might disappear by 2027. But something even more immediate than the disappearance of jobs is happening: those at companies that provide tokens will have their usage curves tracked and displayed. Those at companies that do not have to buy their own tickets to enter the track. No matter whether the ticket is provided by the company or paid for out of pocket, no one has ever answered them: what exactly do you get for winning the leaderboard? No one knows where this track ends, but no one dares to stop first.
This article is from the WeChat Official Account "Outside the Singularity", authored by wiwi, and published by 36Kr with authorization.