Jul 3, 2026
Token Legends:
What Meta’s AI Leaderboard Reveals About Measuring the Wrong Thing
Reading Time:
8 Minutes
Category:
AI in Education
Metrics, AI fluency, token economics
For a few strange months in 2026, some of the most talented engineers on earth competed for a title that did not exist a year earlier: Token Legend. On an internal Meta leaderboard called Claudeonomics, named after Anthropic’s Claude, more than 85,000 employees could watch their consumption of AI tokens ranked in real time. Session Immortal. Cache Wizard. Model Connoisseur. The vocabulary of a video game, grafted onto the payroll of a trillion-dollar company.
Then the bill arrived. In a single thirty-day window, Meta employees burned through 73.7 trillion tokens. An internal memo to roughly 6,000 staff warned that internal AI use alone was on track to cost billions of dollars in 2026. The leaderboard came down. A monitoring platform called AI Gateway is going up, formal token budgets arrive in 2027, and engineers are being steered from external tools toward Meta’s own coding assistant. Chief Technology Officer Andrew Bosworth wrote the episode’s epitaph in a memo of his own: “All motion is not progress and token usage alone is not a measure of impact of any kind.” The technology press has covered this as a story about cost control. It is not. It is a story about measurement, and it is pointed, whether the industry realizes it or not, directly at education.
A Law Older Than the Leaderboard
In 1975, the economist Charles Goodhart observed that any statistical regularity tends to collapse once pressure is placed upon it for control purposes. Four years later, the psychologist Donald Campbell gave the principle its sharpest social form: the more any quantitative indicator is used for decision-making, the more subject it becomes to corruption pressures, and the more apt it is to distort the very processes it was meant to monitor. The anthropologist Marilyn Strathern later compressed both into a single sentence that every administrator should have framed above the desk: when a measure becomes a target, it ceases to be a good measure.
Why does this happen with such regularity? Because every metric is a proxy. Token counts stood in for productivity the way test scores stand in for learning and citation counts stand in for scholarship. The proxy works only so long as nothing depends on it. The moment rewards attach to the number, rational people optimize the number rather than the thing it once represented, and the connection between the two quietly dies.
At Meta, employees reportedly left AI agents running on idle tasks purely to climb the rankings. That is not a scandal of character. It is Campbell’s law executing exactly as written.
The detail most coverage missed is the most instructive one. Meta’s leadership did not build Claudeonomics. An employee built it, voluntarily, on the company intranet, after performance reviews had made “AI-driven impact” a core expectation and dangled bonuses of up to 200 percent for top performers. This is the part institutions must sit with: you do not need to mandate the game; you only need to make the score matter. The incentive culture will generate its own corruption instruments, cheerfully and from below.
Conspicuous Computation
In 1899, Thorstein Veblen argued in The Theory of the Leisure Class that expenditure functions as status display: we consume visibly in order to be seen consuming. What Silicon Valley now calls tokenmaxxing is Veblen’s logic ported to the machine age, a kind of conspicuous computation.
Nvidia’s Jensen Huang has suggested that engineers should carry annual token budgets approaching half their salaries, and that low token spend by a highly paid engineer should be a cause for concern. Once expenditure itself becomes the signal of being “AI native,” waste stops being an accident and becomes a strategy.
The economics of motivation explain the second-order damage. Bruno Frey’s work on motivation crowding, building on the self-determination research of Deci and Ryan, shows that extrinsic rankings and rewards can crowd out the intrinsic motivations of craft, curiosity, and care. Engineers stopped asking whether the tool helped them build well and started asking whether it raised their number. Reporting from inside Meta describes production incidents traced to careless AI-generated code and throwaway work clustered at the top of the leaderboard. The metric did not merely fail to measure productivity. It actively degraded it.
This Is Not a Meta Problem
If the pathology were confined to one company, it would be an anecdote. It is not. Uber exhausted its entire 2026 AI coding budget in four months after encouraging maximal use and ranking teams on internal leaderboards; it now caps every employee at 1,500 dollars per month per coding tool, and its Chief Operating Officer concedes that the link between token spend and shipped consumer value “is not there yet.” A KPMG survey found that only 26 percent of companies have comprehensive visibility into their AI costs. ServiceNow and Walmart have imposed their own caps.
Nor is it confined to California. In a recent conversation, a senior leader at one of the large technology companies headquartered in the Seattle area told me, unprompted, that the same token wars are unfolding inside his organization: usage dashboards, internal rankings, and engineers who have learned that the number, not the outcome, is what gets seen. Industry reporting has since corroborated that at least one Seattle giant has been running an internal token leaderboard since January.
When the same syndrome appears independently across companies with different cultures, different tools, and different leadership, the diagnosis shifts. Blame the incentive architecture, not the employees.
Education Is Already Building Its Own Leaderboard
The philosopher of education Gert Biesta warned, in Good Education in an Age of Measurement, that institutions under audit pressure end up valuing what they measure rather than measuring what they value. Higher education is now entering precisely the conditions under which that inversion occurs.
Among university students, AI usage has surged to 92 percent, up from 66 percent in 2024, while only 36 percent report receiving formal institutional training in how to use it well. Purdue has approved the first board-mandated AI competency graduation requirement in the United States, effective this fall. Ohio State has built an AI Fluency initiative with uptake metrics as key performance indicators. The California State University system has deployed ChatGPT Edu to more than 460,000 students.
None of these initiatives is wrong in itself. The danger lies in the evidence they will reach for when someone asks whether they are working. When an accreditor, a board, or a ministry asks a university to demonstrate “AI integration,” the cheapest evidence is usage data: logins, prompts, sessions, tokens. Adoption is countable. Formation is not. And the pressure is real: in a January 2026 survey, 90 percent of faculty said they believe AI is weakening student learning, while researchers studying successful institutions note that the distinguishing practice of intentional adopters is that they measure learning-relevant outcomes rather than adoption metrics.
I have argued on this blog that the test of good AI use is what remains when the tool is withdrawn. The tokenization game exposes the institutional mirror of that individual test. A usage dashboard measures the presence of the machine. Education must measure the growth of the person. Those are not the same number, and no dashboard converts one into the other.
An institution that rewards visible AI activity will get visible AI activity, in exactly the way Meta got 73.7 trillion tokens: abundantly, enthusiastically, and increasingly detached from anything worth having.
Three Projections
First, token economics will reach campuses by 2028. Goldman Sachs projects a twenty-four-fold rise in enterprise token consumption by 2030, reaching roughly 120 quadrillion tokens per month industry-wide. Gartner finds that although inference costs will fall by around 90 percent by 2030, agentic AI consumes far more tokens per task and vendors will not fully pass savings through, so institutional AI bills will rise even as unit prices fall. Universities signing enterprise AI agreements today on flat-rate assumptions are underwriting their own Uber moment. Expect token budgets, per-department caps, and AI cost-governance offices inside university administration within two academic years.
Second, AI-fluency requirements will drift into proxy metrics by 2027 to 2028 unless deliberately anchored. The path of least resistance from “requirement” to “evidence” runs straight through usage data. Vendors will market student-facing AI engagement dashboards to provosts, and those dashboards will corrupt exactly as Claudeonomics did: students optimizing visible AI activity for credit, learning to perform fluency rather than possess it. Campbell’s law does not exempt registrars. The alternative is harder and better: assessing fluency through demonstrated performance, with and without the tool, judged by humans who know the student.
Third, verified unaided competence becomes a premium credential by 2030. As AI-assisted output becomes ambient and cheap, its signaling value collapses; what becomes scarce, and therefore valuable, is certified evidence of what a graduate can do with the machine, without the machine, and, crucially, the judgment to know which is called for. Institutions that build assessment regimes capable of certifying all three will command a trust that no usage statistic can purchase. This is not nostalgia for the pre-AI university. It is the market logic of scarcity applied to human formation.
The Question the Dashboard Cannot Ask
Meta can afford its experiment. A few billion dollars of wasted compute is, for a company of that scale, expensive tuition. Universities cannot afford the equivalent, because our waste is not denominated in tokens. It is denominated in graduates.
A university that teaches its students that the score matters more than the work has taught the most corrosive lesson available, and it will have taught it with perfect efficiency, because incentives always teach faster than syllabi.
The dashboards will arrive on our campuses; in some places they already have. When they do, the question every educator, parent, and president must ask is not how much AI our students are using. It is this: are we forming people who could still think, judge, and create if every token on earth went dark tomorrow, or are we merely training them to be legends of a game that measures nothing?
References
3. The Pragmatic Engineer (April 23, 2026). “The Pulse: ‘Tokenmaxxing’ as a weird new trend.”
8. Forbes (December 26, 2025). “7 AI Decisions That Will Define Higher Education in 2026.”
9. Genio (March 2026). “How students are using AI in 2026: A shift from AI adoption to AI agency.”
10. Goodhart, C. A. E. (1975). “Problems of Monetary Management: The U.K. Experience.” Papers in Monetary Economics, Reserve Bank of Australia.
11. Campbell, D. T. (1979). “Assessing the impact of planned social change.” Evaluation and Program Planning, 2(1), 67–90.
12. Strathern, M. (1997). “‘Improving ratings’: audit in the British University system.” European Review, 5(3), 305–321.
13. Veblen, T. (1899). The Theory of the Leisure Class. Macmillan.
14. Biesta, G. (2010). Good Education in an Age of Measurement: Ethics, Politics, Democracy. Paradigm Publishers.
15. Frey, B. S., & Jegen, R. (2001). “Motivation Crowding Theory.” Journal of Economic Surveys, 15(5), 589–611.
For a few strange months in 2026, some of the most talented engineers on earth competed for a title that did not exist a year earlier: Token Legend. On an internal Meta leaderboard called Claudeonomics, named after Anthropic’s Claude, more than 85,000 employees could watch their consumption of AI tokens ranked in real time. Session Immortal. Cache Wizard. Model Connoisseur. The vocabulary of a video game, grafted onto the payroll of a trillion-dollar company.
Then the bill arrived. In a single thirty-day window, Meta employees burned through 73.7 trillion tokens. An internal memo to roughly 6,000 staff warned that internal AI use alone was on track to cost billions of dollars in 2026. The leaderboard came down. A monitoring platform called AI Gateway is going up, formal token budgets arrive in 2027, and engineers are being steered from external tools toward Meta’s own coding assistant. Chief Technology Officer Andrew Bosworth wrote the episode’s epitaph in a memo of his own: “All motion is not progress and token usage alone is not a measure of impact of any kind.”
The technology press has covered this as a story about cost control. It is not. It is a story about measurement, and it is pointed, whether the industry realizes it or not, directly at education.
A Law Older Than the Leaderboard
In 1975, the economist Charles Goodhart observed that any statistical regularity tends to collapse once pressure is placed upon it for control purposes. Four years later, the psychologist Donald Campbell gave the principle its sharpest social form: the more any quantitative indicator is used for decision-making, the more subject it becomes to corruption pressures, and the more apt it is to distort the very processes it was meant to monitor. The anthropologist Marilyn Strathern later compressed both into a single sentence that every administrator should have framed above the desk: when a measure becomes a target, it ceases to be a good measure.
Why does this happen with such regularity? Because every metric is a proxy. Token counts stood in for productivity the way test scores stand in for learning and citation counts stand in for scholarship. The proxy works only so long as nothing depends on it. The moment rewards attach to the number, rational people optimize the number rather than the thing it once represented, and the connection between the two quietly dies.
At Meta, employees reportedly left AI agents running on idle tasks purely to climb the rankings. That is not a scandal of character. It is Campbell’s law executing exactly as written.
The detail most coverage missed is the most instructive one. Meta’s leadership did not build Claudeonomics. An employee built it, voluntarily, on the company intranet, after performance reviews had made “AI-driven impact” a core expectation and dangled bonuses of up to 200 percent for top performers. This is the part institutions must sit with: you do not need to mandate the game; you only need to make the score matter. The incentive culture will generate its own corruption instruments, cheerfully and from below.
Conspicuous Computation
In 1899, Thorstein Veblen argued in The Theory of the Leisure Class that expenditure functions as status display: we consume visibly in order to be seen consuming. What Silicon Valley now calls tokenmaxxing is Veblen’s logic ported to the machine age, a kind of conspicuous computation.
Nvidia’s Jensen Huang has suggested that engineers should carry annual token budgets approaching half their salaries, and that low token spend by a highly paid engineer should be a cause for concern. Once expenditure itself becomes the signal of being “AI native,” waste stops being an accident and becomes a strategy.
The economics of motivation explain the second-order damage. Bruno Frey’s work on motivation crowding, building on the self-determination research of Deci and Ryan, shows that extrinsic rankings and rewards can crowd out the intrinsic motivations of craft, curiosity, and care. Engineers stopped asking whether the tool helped them build well and started asking whether it raised their number. Reporting from inside Meta describes production incidents traced to careless AI-generated code and throwaway work clustered at the top of the leaderboard. The metric did not merely fail to measure productivity. It actively degraded it.
This Is Not a Meta Problem
If the pathology were confined to one company, it would be an anecdote. It is not. Uber exhausted its entire 2026 AI coding budget in four months after encouraging maximal use and ranking teams on internal leaderboards; it now caps every employee at 1,500 dollars per month per coding tool, and its Chief Operating Officer concedes that the link between token spend and shipped consumer value “is not there yet.” A KPMG survey found that only 26 percent of companies have comprehensive visibility into their AI costs. ServiceNow and Walmart have imposed their own caps.
Nor is it confined to California. In a recent conversation, a senior leader at one of the large technology companies headquartered in the Seattle area told me, unprompted, that the same token wars are unfolding inside his organization: usage dashboards, internal rankings, and engineers who have learned that the number, not the outcome, is what gets seen. Industry reporting has since corroborated that at least one Seattle giant has been running an internal token leaderboard since January.
When the same syndrome appears independently across companies with different cultures, different tools, and different leadership, the diagnosis shifts. Blame the incentive architecture, not the employees.
Education Is Already Building Its Own Leaderboard
The philosopher of education Gert Biesta warned, in Good Education in an Age of Measurement, that institutions under audit pressure end up valuing what they measure rather than measuring what they value. Higher education is now entering precisely the conditions under which that inversion occurs.
Among university students, AI usage has surged to 92 percent, up from 66 percent in 2024, while only 36 percent report receiving formal institutional training in how to use it well. Purdue has approved the first board-mandated AI competency graduation requirement in the United States, effective this fall. Ohio State has built an AI Fluency initiative with uptake metrics as key performance indicators. The California State University system has deployed ChatGPT Edu to more than 460,000 students.
None of these initiatives is wrong in itself. The danger lies in the evidence they will reach for when someone asks whether they are working. When an accreditor, a board, or a ministry asks a university to demonstrate “AI integration,” the cheapest evidence is usage data: logins, prompts, sessions, tokens. Adoption is countable. Formation is not. And the pressure is real: in a January 2026 survey, 90 percent of faculty said they believe AI is weakening student learning, while researchers studying successful institutions note that the distinguishing practice of intentional adopters is that they measure learning-relevant outcomes rather than adoption metrics.
I have argued on this blog that the test of good AI use is what remains when the tool is withdrawn. The tokenization game exposes the institutional mirror of that individual test. A usage dashboard measures the presence of the machine. Education must measure the growth of the person. Those are not the same number, and no dashboard converts one into the other.
An institution that rewards visible AI activity will get visible AI activity, in exactly the way Meta got 73.7 trillion tokens: abundantly, enthusiastically, and increasingly detached from anything worth having.
Three Projections
First, token economics will reach campuses by 2028. Goldman Sachs projects a twenty-four-fold rise in enterprise token consumption by 2030, reaching roughly 120 quadrillion tokens per month industry-wide. Gartner finds that although inference costs will fall by around 90 percent by 2030, agentic AI consumes far more tokens per task and vendors will not fully pass savings through, so institutional AI bills will rise even as unit prices fall. Universities signing enterprise AI agreements today on flat-rate assumptions are underwriting their own Uber moment. Expect token budgets, per-department caps, and AI cost-governance offices inside university administration within two academic years.
Second, AI-fluency requirements will drift into proxy metrics by 2027 to 2028 unless deliberately anchored. The path of least resistance from “requirement” to “evidence” runs straight through usage data. Vendors will market student-facing AI engagement dashboards to provosts, and those dashboards will corrupt exactly as Claudeonomics did: students optimizing visible AI activity for credit, learning to perform fluency rather than possess it. Campbell’s law does not exempt registrars. The alternative is harder and better: assessing fluency through demonstrated performance, with and without the tool, judged by humans who know the student.
Third, verified unaided competence becomes a premium credential by 2030. As AI-assisted output becomes ambient and cheap, its signaling value collapses; what becomes scarce, and therefore valuable, is certified evidence of what a graduate can do with the machine, without the machine, and, crucially, the judgment to know which is called for. Institutions that build assessment regimes capable of certifying all three will command a trust that no usage statistic can purchase. This is not nostalgia for the pre-AI university. It is the market logic of scarcity applied to human formation.
The Question the Dashboard Cannot Ask
Meta can afford its experiment. A few billion dollars of wasted compute is, for a company of that scale, expensive tuition. Universities cannot afford the equivalent, because our waste is not denominated in tokens. It is denominated in graduates.
A university that teaches its students that the score matters more than the work has taught the most corrosive lesson available, and it will have taught it with perfect efficiency, because incentives always teach faster than syllabi.
The dashboards will arrive on our campuses; in some places they already have. When they do, the question every educator, parent, and president must ask is not how much AI our students are using. It is this: are we forming people who could still think, judge, and create if every token on earth went dark tomorrow, or are we merely training them to be legends of a game that measures nothing?
References
3. The Pragmatic Engineer (April 23, 2026). “The Pulse: ‘Tokenmaxxing’ as a weird new trend.”
8. Forbes (December 26, 2025). “7 AI Decisions That Will Define Higher Education in 2026.”
9. Genio (March 2026). “How students are using AI in 2026: A shift from AI adoption to AI agency.”
10. Goodhart, C. A. E. (1975). “Problems of Monetary Management: The U.K. Experience.” Papers in Monetary Economics, Reserve Bank of Australia.
11. Campbell, D. T. (1979). “Assessing the impact of planned social change.” Evaluation and Program Planning, 2(1), 67–90.
12. Strathern, M. (1997). “‘Improving ratings’: audit in the British University system.” European Review, 5(3), 305–321.
13. Veblen, T. (1899). The Theory of the Leisure Class. Macmillan.
14. Biesta, G. (2010). Good Education in an Age of Measurement: Ethics, Politics, Democracy. Paradigm Publishers.
15. Frey, B. S., & Jegen, R. (2001). “Motivation Crowding Theory.” Journal of Economic Surveys, 15(5), 589–611.
