Meta Dropped AI Usage From Engineer Reviews After Costs Headed Toward Billions

Meta tied AI tool usage to engineer performance reviews and watched internal AI costs head toward billions of dollars. The metric is gone.

What Happened

Meta built a performance evaluation system that included AI token consumption as a criterion. Engineers discovered that consuming more tokens improved their standings on internal leaderboards — without any requirement that the consumption produce useful work. The behavior was dubbed “tokenmaxxing”: bulk token consumption purely to score higher on the metric.

The consequence was direct. Executives Maher Saba and Santosh Janardhan noted that internal AI costs were “heading toward billions in costs in 2026.” The engineering org had built exactly the incentive structure it designed — optimize for the measured metric — and the measured metric had nothing to do with software quality.

Meta announced it will remove AI usage from engineer performance evaluations. Future assessments will focus on “the quality, speed, and complexity of the work” instead. Starting in 2027, Meta plans to implement usage budgets and a central dashboard to control costs.

Why It Matters

The tokenmaxxing outcome is a textbook Goodhart’s Law problem: when a measure becomes a target, it ceases to be a good measure. Token consumption is a reasonable proxy for AI tool adoption — but adoption is not the same as productivity, and a performance system cannot distinguish between a developer who uses Claude to write a production feature and one who pumps tokens into a benchmark run.

The more significant problem is what the cost trajectory reveals. Meta is one of the largest AI consumers on the planet, with access to internal models, the full Llama family, and every commercial API. If internal AI consumption at Meta is “heading toward billions” on a per-year basis, the token economics that looked manageable at 2024 usage levels are now a budget category that demands governance.

The Pattern

Stack Archive has tracked AI token economics across three angles in the last 30 days: the 14x token self-consumption spike on OpenRouter as agentic systems consume their own context, the FinOps shift from cloud bills to token bills, and GitHub Copilot’s move to consumption-based billing. Meta’s tokenmaxxing case is the same dynamic at the organizational measurement layer: token consumption is now a quantity large enough to require cost controls, budgets, and governance — the same infrastructure that cloud spend required in 2018.

What This Means for Builders

Any organization measuring AI adoption through usage metrics faces the same incentive structure Meta built. If tokens consumed appear on a performance dashboard alongside outcome metrics, the token metric will be optimized independently. The outcome metrics may not follow.

The governance failure here is not about AI — it is about measurement design. Token consumption is an input metric. Shipping reliable software, reducing incident rate, and accelerating review cycles are output metrics. Tying compensation or performance standing to an input metric in any domain produces the same behavior: people optimize the input, not the output.

The 2027 budget and dashboard plan addresses the cost problem. What it does not address is the underlying measurement question: how do you credit an engineer for AI leverage when the productivity gain is real but the token footprint is invisible to current performance systems?

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.