Meta drops AI token counts from performance reviews

Meta has removed AI usage as a performance-review criterion for engineers. Per The Information's reporting, staff were told that AI adoption dashboards and token counts will not be used to evaluate impact. The revised guidance instead frames the outcomes it wants as ones that "can be supported by AI or other means." Which is the sentence you write after you've learned something expensive.

What actually happened

The practice had a name — tokenmaxxing — after a March New York Times column described Meta and OpenAI pushing employees to run up AI usage. Gizmodo's write-up covers the arc: internal leaderboards, heavy users rewarded, light users chastened, adoption turned into a scoreboard. Then a June Financial Times report said Google was throttling Meta's AI tokens, which put a hard ceiling on a metric people were being graded on.

The mixed message is the fun part. Gizmodo notes Meta is simultaneously pushing an internal agentic platform called Hatch to employees — a tool that burns considerably more tokens than a chatbot or a coding assistant. So: stop counting tokens, please use the thing that consumes them. Both instructions can be right, which is exactly why the metric was wrong.

Why this matters for your business

You cannot make adoption a target and also a measurement. The moment a token count appears on a review form, it stops describing behavior and starts producing it. Engineers are extremely good at optimizing whatever is on the dashboard. Meta ran this experiment at a scale you never will, and the result is available for free: the number went up and it did not mean what they wanted it to mean.

Measure the output, not the input. Cycle time from ticket to merged. Defect escape rate. Reviewer time per PR. Support tickets resolved without escalation. Every one of those moves if agents are actually helping, and none of them can be gamed by opening more sessions. We looked at the same input-versus-output gap in Meta's own coding-agent numbers — 220% more code, 36% of it shipped.

Supply is not guaranteed, so don't build a process that assumes it. A vendor throttled a hyperscaler's token allocation mid-cycle. If that can happen to Meta, your rate limit is a business risk too. Cap spend per project, know which parts of your workflow degrade gracefully without an agent, and keep a second provider configured.

If you're a small studio, the honest version of this is cheaper. Pick two or three agent-assisted workflows, run them for a quarter, and compare the delivery metrics you already track against the quarter before. That's the whole program. You don't need a leaderboard, and if you build one you'll get a leaderboard's answer. Related: a third of organizations now skip a software purchase and build internally — that decision deserves outcome metrics too.

Sources: The Information, Gizmodo.