Should you rank engineers by AI token use?
Short answer
Usually not. A ranking by token volume rewards spending, not results. Meta's employee-built “Claudeonomics” board came down in April 2026, and Amazon scrapped its usage leaderboard after people gamed it, which raised compute costs. Track weekly active use, capacity used per seat and outcomes instead, and treat heavy users as teachers, not targets.
On this page
What happened at Meta
In late 2025, Meta told staff that AI-driven impact would be a “core expectation” from 2026, while leaving individual AI usage metrics out of 2025 reviews.1 By April 2026, an employee had built a dashboard called “Claudeonomics” that ranked the top 250 token users out of more than 85,000 employees.
| Item | Reported figure |
|---|---|
| People ranked | Top 250 of more than 85,000 employees |
| Total use on the dashboard, 30 days | More than 60 trillion tokens |
| Top-ranked individual | Averaged 281 billion tokens |
| How long it lasted after the news | Went down two days after the news broke |
The dashboard's closing message, as quoted by Fortune
“due to data from this dashboard being shared externally, we've made the decision to shutter Claudeonomics for now”
Fortune, April 9, 2026 · Checked .
Meta said the employee took it down at their own discretion and that Meta did not ask for it.2 By June 2026, Meta was reported to be reining in staff spending on AI tools.3
What happened at Amazon
Amazon scrapped an internal leaderboard that ranked staff by AI usage after people gamed it, which sent compute costs up.3 The same report describes Uber capping employees at $1,500 a month per tool after using its 2026 AI coding budget by April, and some Microsoft staff spending $500 to $2,000 a month on Claude Code tokens alone.3
Why token rankings fail
- Goodhart's law. Once a measure becomes a target, it stops measuring what it did. Tokens are easy to produce on purpose: a longer prompt, a bigger context, an agent left looping. Amazon's experience is the predictable result.3
- Tokens measure input, not output. A session that rereads a large codebase all day can top the chart and ship nothing. A careful engineer who keeps context small ranks low for doing the efficient thing.
- Quality is not counted. In Stack Overflow's 2025 survey, 66% of developers named AI answers that are almost right, but not quite, as their top frustration.4 More tokens can mean more of that, and more review work for someone else.
- Cost is real. A contest to spend is a budget problem by design. See How much should you budget for AI tokens per developer?
- The data leaks. Meta's board closed because its data was shared outside the company.2 Per-person usage is sensitive by nature.
DORA's 2025 report describes AI as “a mirror and a multiplier”: it boosts efficiency in cohesive organizations and highlights weaknesses in fragmented ones.5 A leaderboard shows how much was multiplied, not whether it was the right thing.
The case for some visibility
To be fair to leaderboards: seeing who uses the tools most is useful for finding people who can show others how. Anthropic's own Claude Code analytics for Team and Enterprise includes a leaderboard of the top 10 users by contribution, counted in pull requests or lines of code with Claude Code, not tokens. Anthropic's docs suggest using it to find people who can share techniques and help onboard others.6 That is a teaching use, private to admins, and based on merged work. It is a long way from a company-wide token contest.
What to track instead
| Instead of | Track | Why it holds up |
|---|---|---|
| Tokens per engineer, ranked | Share of seats active each week | Measures reach. Hard to fake without doing the work. |
| Top spender | Share of weekly limit used per seat, at the reset | Finds idle capacity and seats that are too small |
| Most sessions | Time agents wait on a person | Points at process bottlenecks |
| Lines generated | Merged PRs, lead time, change failure rate | Counts work that shipped and stayed shipped |
Details and scripts: How do you measure Claude Code adoption across a team? and Are you using the Claude capacity you pay for?
If you show a leaderboard anyway
Rank outcomes, not tokens
Merged pull requests with AI help, not volume consumed.
Keep it private or team-level
Admins and team leads, not a company-wide wall.
Keep it out of reviews
Meta kept individual usage metrics out of its 2025 reviews.1 Review impact, which a person can explain, not a count, which a person can inflate.
Put a budget next to it
Every usage number shown should sit beside its cost.
Common questions
What was Meta's Claudeonomics leaderboard?
An employee-built dashboard at Meta that ranked the top 250 token users out of more than 85,000 employees. Total use on it passed 60 trillion tokens in 30 days. It went down in April 2026, two days after the news broke.
Why did Amazon drop its AI usage leaderboard?
According to The Next Web in June 2026, Amazon scrapped the leaderboard after people gamed it, which sent compute costs up.
Does Claude Code have a built-in leaderboard?
Team and Enterprise analytics include a leaderboard of the top 10 users by contribution, measured in pull requests or lines of code with Claude Code. Anthropic's docs suggest using it to find people who can help others get started.
Is it wrong to look at per-person AI usage at all?
No. Per-person numbers help with budgets, seat sizes and spotting someone who is stuck. The problem starts when the number is ranked publicly or used as a target.
Sources
- Meta to formally review employees' AI performance from 2026HR Grapevine, Nov 17, 2025Secondary report
- A Meta employee created a dashboard so coworkers can compete to be the company's No. 1 AI token userFortune, Apr 9, 2026Checked
- The tokenmaxxing era is over. Now companies are 'tokenminimizing'The Next Web, Jun 17, 2026Checked
- 2025 Developer Survey: AIStack Overflow, 2025Checked
- How are developers using AI? Inside our 2025 DORA reportGoogle, Sep 23, 2025Checked
- Track team usage with analyticsClaude Code docs, undatedChecked