An internal leaderboard logging employee AI use — nicknamed 'Claudeonomics,' tracking how much workers used Anthropic's Claude and other AI models — clocked 73.7 trillion tokens (the small text chunks AI models process) in just over 30 days as Meta employees raced to top a usage…
Meta built an internal leaderboard called "Claudeonomics" to track how much value employees were getting from AI. Instead, the board tracked something else: 73.7 trillion tokens, the small text chunks AI models process, in just over 30 days, generated by employees racing to top a ranking that rewarded volume over usefulness. The race itself is now the story, because it shows what happens when a company makes AI use a performance metric and then tries to govern it.
The leaderboard sat inside a broader push that, earlier in 2026, had framed AI usage as a "core expectation" in performance reviews, according to The Information, as reported by The Decoder. Employees at Meta, the parent of Facebook, Instagram, and WhatsApp, responded by inflating their personal AI consumption, a practice the company has internally labeled "tokenmaxxing." The 73.7 trillion tokens logged on the Claudeonomics board were the visible score of that behavior.
The cost has grown large enough that Meta circulated an internal memo to roughly 6,000 employees flagging an "exponential increase" in internal AI usage, with that usage on track to cost the company billions of dollars by the end of 2026, per The Information, as reported by The Decoder. The Information's underlying reporting is paywalled; only the secondary summary is in hand this turn.
Meta is now responding with structure. Starting in 2027, the company plans tighter token management: per-team budgets, allocations, and dedicated tools. An internal engineering team has built a central "AI Gateway" dashboard that already tracks AI usage and spend, and automatic alerts for unusual cost spikes are slated to come next. The stated goal is to give employees and managers visibility they did not previously have.
The harder problem is measurement. CTO Andrew Bosworth has publicly argued that "token usage alone is not a measure of impact," a position that openly contradicts the earlier "core expectation" memo. That contradiction is the spine of the case: a company that builds frontier AI cannot reliably tell, from inside its own systems, which AI use is productive and which is performance.
Meta is also steering employees away from third-party tools, including Anthropic's Claude, toward its in-house coding assistant MetaCode, while keeping some external models available. The framing is consistency and cost control. The effect is the same as the leaderboard's: a single, company-defined definition of good AI use, imposed top-down.
The Claudeonomics episode generalizes beyond Meta. Any organization that tells workers to "use more AI," and then measures them on usage, will get usage. The harder question, which the Meta case sharpens, is what good AI productivity measurement actually looks like when even the companies building the systems cannot separate productive use from performative use. Token counts and leaderboards are easy to produce. Useful measurement is the part Meta, like every other employer issuing AI mandates, has not yet built.