A Meta employee built a dashboard tracking AI token consumption across 85,000 workers. Sixty trillion tokens in thirty days. Gamified badges. Shut down within days when the data leaked. The most expensive demonstration of Goodhart's Law in corporate history.
A Meta employee built an internal dashboard called Claudeonomics — named after Anthropic's Claude, the model that had become Meta's primary coding tool. The dashboard tracked AI token consumption across the company's more than eighty-five thousand employees, ranked the top two hundred and fifty users, and awarded gamified badges: bronze, silver, gold, platinum, jade. Achievement titles included Cache Wizard, Model Connoisseur, Token Legend, and Session Immortal. The top individual consumed two hundred and eighty-one billion tokens in thirty days. The company collectively burned through sixty trillion.
The dashboard was shut down within days of its existence becoming public. The message left in its place read: "It was meant to be a fun way for people to look at tokens, but due to data from this dashboard being shared externally, we've made the decision to shutter Claudeonomics for now."
The stated reason for killing it was a data leak. The actual lesson is older than software.
The Mandate
Claudeonomics did not emerge in a vacuum. In late 2025, Janelle Gale, Meta's head of people, announced that "AI-driven impact" would become a core performance expectation starting in 2026. In January, the company overhauled its performance review system with bonuses of up to two hundred percent for top performers. Zuckerberg instructed engineering teams to rewrite the existing codebase to make it parsable by AI. Heavy token usage became the leading indicator that engineers were doing this work.
The leaderboard was a bottom-up response to a top-down mandate. An employee built the measurement tool the company implicitly demanded. And because what gets measured gets optimized, eighty-five thousand people immediately began optimizing for the metric that was now tied to their compensation.
Two hundred and eighty-one billion tokens in thirty days from a single user averages nine point four billion tokens per day. For context, a typical Claude conversation uses one thousand to ten thousand tokens. The number is not a measure of productivity. It is a measure of how hard someone pressed the accelerator.
The Pattern
In 2016, Wells Fargo acknowledged that employees had created three point five million fraudulent bank and credit card accounts. The cause was an internal sales mandate — "eight is great" — that set a target of eight Wells Fargo products per customer household. Employees who could not meet the quota through legitimate sales created accounts without customer knowledge or consent. The bank paid three billion dollars in criminal and civil penalties. The CEO resigned. The metric had become the product.
Amazon issued a similar mandate in late 2025. An internal memo signed by two senior vice presidents established Kiro, Amazon's AI coding assistant, as the standard tool across the company, with an eighty percent weekly usage target. Adoption was tracked via management dashboards. Exceptions required VP-level approval. Approximately fifteen hundred engineers signed an internal petition against the mandate, arguing it prioritized corporate product adoption over engineering quality. Within three months, an AI agent deleted a production environment and two outages wiped out six point three million orders.
Meta's Claudeonomics sits at the intersection of these two patterns. Like Wells Fargo, the metric created selection pressure for volume over value — tokens consumed rather than problems solved. Like Amazon, the mandate turned an engineering tool into a compliance instrument. The distinguishing feature was the gamification. Wells Fargo had quotas. Amazon had dashboards. Meta had a leaderboard with jade badges and achievement titles. The progression reveals something about how organizations absorb new mandates: first compliance, then competition, then spectacle.
Goodhart's Law at Organizational Scale
Charles Goodhart's original observation was narrow: when a measure becomes a target, it ceases to be a good measure. The British economist was describing monetary policy in 1975. The principle has since been observed in education (teaching to the test), healthcare (hospitals gaming readmission metrics), policing (arrest quotas distorting crime statistics), and software engineering (lines of code as a productivity measure).
Token consumption is lines of code for the AI era. Both are input metrics masquerading as output metrics. Both are trivially gameable. Both reward volume over value. And both create a specific organizational failure mode: the appearance of adoption without the substance of capability.
The sixty trillion tokens Meta consumed in thirty days are worth an estimated nine billion dollars at public API pricing. Meta does not pay public rates — it has enterprise agreements and runs many models internally. But the number illustrates the scale of what was being optimized. Even at deeply discounted rates, the token consumption represents a meaningful cost. The question Claudeonomics could not answer is what that cost produced.
What the Leaderboard Measures
The dashboard was killed because the data leaked. But the data was not the problem. The problem was what the data revealed: that a company which mandated AI adoption by metric had purchased compliance, not capability.
A Cache Wizard who optimizes token consumption may be writing better code, or may be running the same prompt repeatedly with minor variations. A Token Legend who averages nine billion tokens per day may be building transformative internal tools, or may be feeding entire codebases into context windows to inflate a number tied to a two-hundred-percent bonus. The leaderboard cannot distinguish between these. That is not a flaw in the leaderboard. It is a flaw in measuring adoption by consumption.
The deeper pattern is organizational. When leadership mandates tool usage and ties it to compensation, the mandate creates two populations: those who use the tool because it makes their work better, and those who use the tool because the metric requires it. The first group would have adopted without the mandate. The second group produces the volume that makes the dashboard impressive and the capability gains that make the dashboard misleading.
Wells Fargo's cross-selling mandate created the same bifurcation. Some employees genuinely helped customers open accounts they needed. Others created accounts that existed only to satisfy the metric. The scandal was not that everyone cheated. It was that the system could not tell the difference between the employees who did and the employees who didn't — and the metric rewarded both equally.
Meta shut down the leaderboard. It has not shut down the mandate. AI-driven impact remains a core performance expectation. The two-hundred-percent bonus structure remains in place. The measurement instrument is gone. The selection pressure is not.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)