Meta burned 60 trillion tokens in 30 days. An engineer built an internal leaderboard called "Claudeonomics" that ranked 85,000 employees by AI token consumption. The top individual averaged 281 billion tokens per day. People left agents running overnight on busywork to climb the ranking. Meta killed the leaderboard after The Information reported on it and the backlash hit.
This is not a Meta problem. This is a pattern across every company measuring AI developer productivity by usage volume.
Salesforce set minimum spend targets: $100 for Claude Code, $70 for Cursor. Engineers at Microsoft admitted to deliberately inflating metrics, including asking AI to process documentation slowly or prototyping features they planned to discard. Gergely Orosz called it what it is: the new lines-of-code metric. Easily gamed. Inversely correlated with value.
Why AI coding tool acceptance rates are misleading
Management sees 80-90% code acceptance rates and concludes the tools are working. They are not looking at what happens the following week.
Real-world acceptance rates drop to 10-30% when you account for code that gets rewritten, reverted, or quietly replaced. The data:
Code churn increased 39% in projects with heavy AI tool adoption (TechCrunch, 2026)
One study found churn up 861% under the highest adoption levels
Engineers achieved 2x throughput at 10x the token cost
If half your PRs need rewriting within a week, you have not doubled output. You have doubled churn and called it progress.
The real cost of measuring AI tool usage as a KPI
Money. 60 trillion tokens at standard pricing would be ~$900 million. Most companies are not Meta, but the pattern scales: teams burning $50K-$100K/month with no clear connection between spend and shipped features.
Judgment decay. When the incentive is "use more AI," the rational response is to use AI on everything. Including things that do not benefit from it.
Culture rot. When token usage becomes a proxy for developer productivity:
Engineers who think carefully before prompting look slow
Engineers who reject bad output and rewrite manually look unproductive
Engineers who read the codebase before generating anything look like they are not using the tools
The incentive structure punishes exactly the behavior you need: judgment.
AI-assisted code produces 1.7x more major issues when developers do not validate output (Second Talent, 2025). METR found experienced developers were 19% slower with AI because verification takes time (METR.org, 2025). The engineers who slow down are the good ones. Tokenmaxxing punishes them for it.
How to measure AI developer productivity instead
Shopify got closer to reasonable. They renamed their "leaderboard" to "usage dashboard," added circuit breakers, manager oversight of high spenders, and explicitly told engineers that high usage is not a signal of high performance.
The principle: measure output, not input. Token consumption is an input metric. Code that ships, stays shipped, and does not create follow-up work is an output metric.
If you are a CTO or VP Eng:
Check code churn rate pre- vs post-AI adoption. If churn went up, your acceptance rates are lying
Look at the ratio of PRs that survive a full sprint without modification
Stop measuring token spend as a KPI. It is a cost center, not a performance indicator
The engineers who produce the best work with AI prompt carefully, validate critically, and know when to close the tool and think for themselves.
At Fairground, the AI Coding Screener captures how candidates actually work with AI, not how many tokens they burn. The scorecard measures judgment: did they validate output, catch model errors, know when to stop? That is the signal that predicts quality code, not volume. Start free, 100 credits.
Related: The Engineering Job Description Is Different Now
Get started with Fairground in just few mins.
Plug and Play. Works well with your existing ATS.
100 Free Credits
