Token leaderboards turned AI usage into a target. The better question is what changed because AI was used.
A follow-up to: The Economics Angle: Cheaper AI, Bigger and Unpredictable Bills
In my last article, I wrote about a paradox that's quietly unnerving finance teams everywhere: per-token costs keep falling, yet enterprise AI bills keep climbing. The culprit wasn't the price per unit; it was the relentless, often unchecked expansion of usage.
What I didn't predict was how fast companies would turn that usage into a performance metric.
Meet tokenmaxxing.
Earlier this year, a Meta employee independently built an internal leaderboard called "Claudeonomics", tracking how many AI tokens each of Meta's 85,000+ employees consumed. The top 250 were ranked. Titles were awarded: "Token Legend." "Cache Wizard." "Model Connoisseur."
In a single 30-day window, total usage across that dashboard exceeded 60 trillion tokens. The highest individual averaged 281 billion tokens per month. For context: neither Mark Zuckerberg nor CTO Andrew Bosworth made the top 250.
The dashboard was shut down two days after The Information reported on it, but not, it turns out, because Meta ordered it. The employee who built it pulled it themselves, citing concerns about internal data leaking externally. Meta later clarified they hadn't requested its removal.
Around the same time, Amazon was running its own version of this experiment. Its internal AI platform, "MeshClaw", lets employees spin up agents to handle tasks: code deployments, email triage, Slack interactions. Amazon set a target for over 80% of its developers to actively use AI tools each week, and started tracking token consumption on what the company calls "dashboards" (and what employees called leaderboards).
The Financial Times reported what followed: some workers were using MeshClaw (at Amazon) to automate entirely unnecessary tasks, not to get work done, but to inflate their token numbers. "There is just so much pressure to use these tools," one Amazon employee told the FT. "Some people are using MeshClaw simply to maximise their token usage."
Officially, Amazon said the numbers wouldn't affect performance reviews. Multiple employees said they believed managers were watching anyway.
This is Goodhart's Law, arrived at lightning speed: when a measure becomes a target, it ceases to be a good measure.
Here's the problem no leaderboard accounts for: those tokens cost real money whether or not they produce real output.
This is the throughline from my previous piece: cheaper unit costs and more usage don't cancel each other out. They multiply. And if that usage is being gamified rather than directed, the ROI math gets even harder to defend.
This is where it gets genuinely interesting, and where I think the tokenmaxxing instinct gets the causality exactly backwards.
The landmark NBER study on generative AI at work (Brynjolfsson, Li, and Raymond, 2023) studied 5,179 customer support agents at a Fortune 500 firm and found that AI tools increased productivity by 14% on average. But that average conceals the real story.
The AI wasn't conferring an advantage on the people already most capable of doing the work. It was closing the gap between them and everyone else, by encoding and distributing the tacit knowledge that top performers had developed over years.
That finding has a sharp implication for how we think about measurement. The most valuable thing the AI was doing was almost invisible by a token-usage metric. It was transferring expertise, not consuming compute. A senior agent who ignores the AI's suggestion and resolves the issue faster generates fewer tokens than a junior agent using it heavily to work through the same problem. Which output is more valuable?
The framing of tokenmaxxing gets this entirely wrong. It treats AI usage volume as a proxy for AI-driven value. But if the research is right, the most efficient outcomes may actually show up as lower token usage relative to output, because the AI has helped someone get better at the work, not just get through more of it.
Fortune recently declared that tokenmaxxing is over. The companies that chased usage as an end in itself are pulling back: limiting access to advanced models, setting token budgets, installing consumption controls. Microsoft cancelled Claude Code subscriptions across several product divisions. Meta took down the leaderboard.
The Forbes piece by Tim Keary put it plainly: tokenmaxxing is out, valuemaxxing is in.
But what does "value" actually mean in this context?
It's not a softer version of the same metric. It's a different question entirely. Instead of asking how much AI is being used, the better question is: what changed because it was used? Did a workflow shorten? Did a decision sharpen? Did someone who was six months away from being effective get there in two?
These are harder to measure. They don't produce clean leaderboards. But they're the questions that map onto outcomes finance teams actually care about.
If you're building or managing AI adoption right now, here's the reframe worth sitting with:
Token volume is an input metric. Value is an output metric. You need to decide which one you're actually managing.
Designing incentives around token consumption is like managing a factory by measuring how much electricity the machines use, rather than what they produce. The correlation exists, but it's loose, gameable, and tells you almost nothing about whether the right things are getting made.
The companies that come out ahead in the next 18 months won't be the ones that burned the most tokens. They'll be the ones that figured out where AI genuinely closes a gap (in skill, in speed, in judgment) and built their measurement frameworks around that.
Token maxxing got us into the room. The map is a different project.
Sources: Fortune (April 2026), Financial Times (May 2026), GitHub Blog (April 2026), Forbes / Tim Keary (2026), Brynjolfsson, Li & Raymond, "Generative AI at Work," NBER Working Paper 31161 (2023, updated 2026). Originally published on LinkedIn.