Tokenmaxxing: Why AI Usage Leaderboards Backfire

Reading Time: 6 minutes

If your company starts counting how much AI you use, what exactly is it measuring: your work, or your willingness to play along?

🪙 The Token Diaries, Episode 3 of 8 — Week one mapped the seven meanings of a token. Episode 1 introduced the employer’s lens, and Episode 2 showed the developer’s ledger. Today we go deep on what happens when usage itself becomes the scoreboard.

🏁 Tokenmaxxing: The Scoreboard Arrives

There is now a word for it: tokenmaxxing. It means pushing your AI usage as high as possible because someone is counting.

The idea is simple, and so is the trap. Management wants proof that employees are adopting AI. Tokens are easy to count. So the count becomes a dashboard, the dashboard becomes a leaderboard, and the leaderboard becomes a quiet instruction: use more.

📌 Real-world case: CIO.com reported in June 2026 that Amazon, JPMorgan, Meta and Disney had reportedly set up AI usage leaderboards, in some cases pushing workers to run up large bills. One example it cites, attributed to Business Insider: a single Disney employee interacted with Claude 460,000 times in nine days. (CIO.com)

Pause on that number. 460,000 interactions in nine days is a machine’s workload. A human did not do 51,000 meaningful things a day. That is what a number looks like once it becomes a target.

🪙 Token Truth: A count that anyone can inflate will eventually be inflated.

🧪 The Old Law Nobody Reads

This is not a new problem. It is a very old one wearing a new badge.

In 1975, the economist Charles Goodhart wrote about UK monetary policy and observed that a statistical regularity tends to collapse once you put pressure on it for control purposes. The version most of us quote came later. In 1997 the anthropologist Marilyn Strathern put it plainly: when a measure becomes a target, it ceases to be a good measure.

Toh phir, the question for any leaderboard is simple. Was the number meant to describe good work, or to become the goal?

📌 Real-world case: Wikipedia’s account of the law points to a pandemic-era example. The UK government set a goal of 100,000 COVID tests a day. The definition first counted tests carried out, then changed to count testing capacity. Researchers noted that useful diagnostic tests were far fewer than the figure reported when the target was declared met. The measure had stopped measuring. (Wikipedia: Goodhart’s law)

🪙 Token Truth: Goodhart did not predict tokenmaxxing. He described it fifty years early.

🏭 The Bill Behind the Board

Here is where it stops being a culture problem and becomes a finance problem. Gamed tokens are real tokens, and real tokens get billed.

By mid-2026 the pullback had started. Fortune’s May 28 column described the sticker shock and noted that many companies were limiting who can use third-party agents built on the most advanced models.

📌 Real-world case: Microsoft’s Experiences + Devices group, which covers Windows, Microsoft 365, Outlook, Teams and Surface, began cancelling most internal Claude Code licences and moving developers to GitHub Copilot CLI, with Claude Code use expected to end by late June 2026. The group’s executive vice president framed it as consolidation. Unnamed sources cited by The Verge said cost mattered too, since the fiscal year ends June 30. (The Decoder, citing The Verge)

📌 Real-world case: Salesforce CEO Marc Benioff said his company’s Anthropic bill would be around $300 million this year, and wished for a “smart router” that sends simpler queries to cheaper models. That is last week’s Episode 2 lesson, voiced by a CEO. Uber’s COO added that token costs were getting “harder to justify” as productivity gains proved hard to link to company-wide results. (Fortune)

Notice the shift. In week one of a leaderboard, burning tokens looks like enthusiasm. In month four, it looks like a budget line someone has to explain.

🪙 Token Truth: Every vanity metric eventually sends an invoice.

⏳ The Life Cycle of a Leaderboard

Line up the 2026 stories and a pattern shows up.

First comes the launch. A mandate or a dashboard arrives, and usage jumps. That looks like success.

Then comes the gaming. People find the cheapest way to move the number, whether that is running agents on trivial tasks or asking the same thing five ways. The chart keeps climbing while the work stays the same.

Then comes the invoice. Finance sees the bill, asks what it bought, and nobody can answer in business terms.

Finally comes the retreat: leaderboards taken down, licences cut, caps introduced. The tools survive, but the scoreboard does not.

The lesson is not to avoid measuring AI. It is to know that a measure which rewards volume will always run through these four stages.

🔄 Update: A Review Criterion That Did Not Survive

Duolingo is a clean example of the retreat stage, with one difference: it never ran a token leaderboard. It put AI use into performance reviews.

📌 Real-world case: In April 2025, CEO Luis von Ahn announced Duolingo would be “AI-first”. Fortune reported the plan: evaluate AI use in annual reviews, add headcount only if a team could not automate more of its work, and gradually stop using contractors for work AI can handle. Public backlash followed. In April 2026 he told the Silicon Valley Girl podcast the company had backtracked: performance should rest on doing the job well, and “I’m not going to force you to do that.” He said employees had asked whether Duolingo wanted AI used for AI’s sake. (Fortune, 24 May 2025; Fortune via Yahoo Finance, 13 Apr 2026)

I could confirm the review criterion was dropped. I could not confirm what happened to the other parts of the 2025 memo, so I am not claiming they changed.

Read it against the lifecycle. The metric started as a signal of adoption, became something employees were asked to show, and then raised the question Goodhart predicts: is this number about good work, or about the number? The fix von Ahn describes is the one this episode ends on: judge outcomes, not use.

🧑‍🏫 Why Smart People Play the Game

Before we blame employees, a fair question: why does this happen to capable people?

Because incentives are loud and nuance is quiet. If using more AI is visible and rewarded, and doing better work is hard to see, people optimise what they can see.

📌 Real-world case: After the Financial Times reported on Amazon employees padding usage, D.A. Davidson analyst Gil Luria told Fortune the trend “doesn’t sound very healthy” and said people do what incentives reward. Rewarding heavier resource use, he argued, produces heavier use. (Fortune)

That is not a character flaw. It is an engineering truth about people. Design the scoreboard badly and you get a game, not a result.

🪙 Token Truth: People do not game metrics because they are dishonest. They game them because the metric asked them to.

🙅 Not Everyone Is Playing

It is worth noting the other side of the trade. Some of the people who sell the tokens are not running a leaderboard of their own.

📌 Real-world case: In remarks from the Bloomberg Tech conference reported on June 4, 2026, Anthropic president Daniela Amodei talked about AI adoption. According to Business Insider’s report, as relayed by LetsDataScience, Anthropic has no internal token-usage leaderboard, though it tracks general use of its products. She hoped AI becomes part of daily work in a way that “feels really good to people.” (LetsDataScience, citing Business Insider)

Read it carefully. She did not argue for or against tokenmaxxing. Adoption you want people to feel good about is a different goal from adoption you force with a ranking.

🪙 Token Truth: Adoption people feel good about beats adoption people are ranked into.

📏 Measure What Shipped

So what should a company count? The CIO.com piece collects advice from practitioners that is worth stealing.

Pair tokens with output. Reward what reaches production, not what was typed or generated.
Track wasted spend. Money spent on work that was rejected, reworked or never shipped is the real leak.
Ask about value, not use. Once people are using AI, ask whether it is for real value or just for the sake of using it.
Look at outcomes. Efficiency, quality and risk reduction beat sheer volume.
Match model to task. Frontier models on simple jobs waste money, which is Episode 2 again.

A token count should sit next to an outcome metric, never replace it.

And if you are the employee, not the manager? Keep a short record of what you shipped using AI and what it saved. When someone finally asks for proof of value, that note beats any dashboard.

Sach batao: if your manager had a dashboard showing only how many tokens you used last week, what would it miss about your work?

🔜 Next Tuesday: Episode 4 — Your Prompts, Their Data

📬 Episode 4 is on its way. Subscribe to get it in your inbox the moment it goes live.

Every prompt is a log entry. This week we asked what companies count. Next week we ask what they can see: skill, surveillance and the employee’s side of the ledger.

Got a story about AI dashboards at your workplace? Drop it in the comments. The best ones may make it into the series.

📚 References

What’s your Reaction?