Tokenmaxxing: How "Use More AI" Became a KPI Engineers Game

The metric ate the mission

Sometime in the last year, a number that used to mean “how much compute this task consumed” quietly got promoted. Token spend stopped being a cost line and became a scoreboard. At Meta, Microsoft, and Salesforce, engineers now have a reason to want their AI usage to go up — not because the work demands it, but because someone upstairs decided that AI adoption could be measured by how many tokens flow through a developer’s account.

Predictably, the tokens started flowing.

The Pragmatic Engineer’s reporting gave the behavior a name: tokenmaxxing. Internal leaderboards rank engineers by consumption. People ask the model redundant questions. They prototype features they have no intention of shipping, purely to move their number. Meta’s employees reportedly generated 60.2 trillion tokens in 30 days — at Anthropic’s list API prices, roughly $900 million of usage. The leaderboard got quietly abolished after the press noticed.

This is Goodhart’s law in its purest form: when a measure becomes a target, it ceases to be a good measure. And it’s worth being precise about what broke, because the failure isn’t AI-specific. It’s a management failure that AI happens to make unusually expensive.

Why token count was always the wrong proxy

The logic behind the KPI is seductive. Leadership believes — probably correctly — that engineers who skillfully use AI tools ship faster. TechCrunch’s reporting puts heavy AI users at roughly twice as productive as light users. So far, so reasonable.

The leap comes next: if good engineers use a lot of AI, then using a lot of AI must make you a good engineer. That’s affirming the consequent, and the tooling launders it into policy. Token count is trivial to measure and impossible to fake your way out of, which makes it irresistible to a VP who needs a slide. The problem is that the same heavy users who are 2x more productive spend roughly 10x more tokens to get there. The correlation between value and consumption is weak and the causation runs the wrong way — but you can only see that if you measure outcomes, and outcomes are hard.

Consider what the metric actually rewards. Suppose you need a one-line null check. The value-maximizing move:

if user is None:
    raise ValueError("user required")

The token-maximizing move is to open an agent, ask it to “analyze the validation strategy across the module,” let it read fifteen files, propose a refactor, and regenerate a test suite — then keep the one-liner. Same shipped code. A few thousand times the spend. Under a usage KPI, the second engineer looks more “AI-native.” Microsoft staff describe exactly this pressure: tokenmax or risk being labeled a laggard.

Salesforce wired the incentive even more directly, surfacing a minimum expected spend — on the order of $100/month — and making the ceiling easy to bypass. Tell people the floor and they’ll stand on it. Tell people there’s no real ceiling and a few will spend $500 a day, or $10,000 in a week.

The bill is not theoretical

The gamification meets reality at the invoice. Uber reportedly burned its entire 2026 AI coding budget by April. Priceline saw Cursor renewal costs jump 4–5x. One company ran up a $500 million Anthropic bill after failing to set limits. The FinOps Foundation’s J.R. Storment says many companies are now 3x over their 2026 token budgets — what he calls an existential crisis in spend management.

There’s a second-order cost that’s easier to miss. AI-generated code that nobody asked for still has to be reviewed, and careless generation produces real incidents — Meta’s spree reportedly caused production SEVs. The bottleneck just moves from writing code to reviewing the firehose of it. You don’t get free throughput; you relocate the constraint to your most senior people and call it progress.

Follow the revenue chart

Here’s the part that should make any engineering leader uncomfortable. The same week this reporting landed, Anthropic confidentially filed for an IPO on the back of annualized revenue crossing $47 billion in May, up from about $9 billion at the end of 2025. Co-founder Daniela Amodei waved off worries about corporate spending cuts by arguing companies are still early in deploying AI well.

She may be right about the long run. But notice the alignment: a customer-side metric that rewards raw consumption maps perfectly onto a vendor-side metric that is raw consumption. Every wasted token in a tokenmaxxing leaderboard is recognized revenue on someone’s pre-IPO chart. That doesn’t require a conspiracy — it just requires that nobody on the buyer’s side is incentivized to ask whether the spend produced anything. A usage KPI guarantees nobody is.

What measuring the right thing looks like

The fix isn’t to ban AI or to cap everyone at a stingy budget. It’s to stop conflating the proxy with the goal.

Shopify offers the template. It started with a leaderboard, recognized the incentive it was creating, and rebranded it to a plain usage dashboard — monitoring for context and safety rather than competition. It added circuit breakers to kill runaway agents and required manual review of the heaviest spenders. The reframe matters: the same data used to rank people corrodes behavior, while used to catch anomalies it surfaces real bugs and real waste.

The broader industry is groping toward the same discipline. A wave of observability startups — Pay-i, Faros AI, plus Datadog and Ramp — is racing to make token spend auditable at “trillions of rows a month” scale, and the Linux Foundation is standing up a Tokenomics Foundation to define metrics like cost-per-intelligence. That’s the FinOps playbook arriving a decade late for a faster-growing cost.

If you run an engineering org, the test is simple. Ask what your AI metric rewards when an honest engineer games it perfectly. If the answer is “they ship the same software and spend ten times more,” you haven’t built a productivity metric. You’ve built a billing funnel — and pointed it at your own budget.