Meta Drops AI Token Usage From Engineer Performance Reviews

Picture an engineering manager opening the internal leaderboard on a Monday morning. Her best engineer, the one who shipped the payments refactor, sits at rank 4,000. The intern who left three agents looping over the weekend is in the top 250, with the title "Cache Wizard" next to his name. That, roughly, is the situation Meta built for itself over the past ten months, and on September 8 The Information reported that the company is dismantling it.

The memo

In an internal note, executives Maher Saba and Santosh Janardhan told engineers that AI dashboards and token counters will no longer count toward performance reviews. Assessments will instead look at the quality, speed and complexity of the work delivered. The change reverses a push that started in November 2025, when Meta declared "AI-driven impact" a core expectation for staff and began measuring it in the most literal way available: how many tokens each person sent to a model.

The measuring tool had a name. "Claudeonomics," a nod to Anthropic's Claude, ranked more than 85,000 employees by monthly token consumption and displayed the top 250 publicly, handing out titles such as "Token Legend" and "Cache Wizard" to the biggest consumers.

How a metric turns into a game

People did what leaderboards ask of them. Engineers ran multiple agent tasks in parallel to pump their numbers, kept loops running that nobody needed, and generally spent compute to look busy. Staff called it "tokenmaxxing." Usage hit 60.2 trillion tokens in a 30-day window this spring and peaked at 73.7 trillion before the leaderboard was pulled.

Meta's CTO Andrew Bosworth had already seen the problem coming. In an April memo he wrote that "nobody should be using AI tools just for the sake of using them," adding that "all motion is not progress and token usage alone is not a measure of impact." It took a bill to make the point stick: according to the same reporting, Meta's internal AI use alone is heading toward billions of dollars in 2026. From 2027 the company will set team-level budgets and route spending through a central dashboard called AI Gateway. Meta is not the only one learning this lesson. Uber and ServiceNow are reported to have exhausted their annual Anthropic budgets within months.

One irony worth recording: the same week the counters left performance reviews, Meta began pushing its new personal agent, Muse, onto its own employees. The incentive to burn tokens is gone; the tools that burn them are multiplying.

The lesson for everyone with a smaller org chart

Most companies rolling out AI right now measure adoption because adoption is easy to measure. Seat activation rates, prompts per employee, the percentage of the team that "uses AI weekly": these numbers fill a slide nicely and say almost nothing about output. Meta ran that experiment with 85,000 people and a nine-figure compute bill, so nobody else needs to.

Three things transfer directly. First, keep AI usage out of individual performance scoring entirely; measure the business outcome the tool was supposed to move, such as proposal turnaround or ticket resolution time. Second, still collect the usage data, but for cost control and learning rather than incentives, so you can answer which team gained what from which tool. Third, do not wait until 2027 to set budgets. On an API-based setup, per-person and per-project spending caps are a day-one configuration, not a program. A counter race cost Meta billions before it was called off. A fifteen-person company can skip the race.

Sources: The Information, The Decoder, AI-RTZ (Michael Parekh)