Abstract. On February 4, 2026, Meta became the first major tech company to formally tie performance reviews to AI usage: engineers would be scored on "AI-driven impact," and workers who leaned hardest on chatbots and agents earned labels like "AI Native," "AI First," and "AI Enabled." Source: Winbuzzer, "Meta Ties Employee Performance Reviews To AI Usage Starting 2026," February 2026. Within two months, an internal leaderboard nicknamed Claudeonomics was ranking roughly 85,000 employees by raw token consumption, with 30-day usage climbing from 60.2 trillion tokens to 73.7 trillion. Source: Tom's Hardware, "AI cost crisis hits tech giants as employee tokenmaxxing backfires," May 2026. Employees started instructing agents to run multiple tasks in parallel purely to inflate their rank. The leaderboard came down within days of leaking. Seven months after it started, the mandate came down too: the week of September 2, 2026, Meta told engineers, in writing, that it "will not use AI adoption dashboards or token counts to evaluate impact," retiring the AI Native/First/Enabled labels for language that says results "can be supported by AI or other means." Source: The Information, "Exclusive: Meta Tells Engineers AI Token Usage Won't Be Part Of Performance Reviews," September 2026. Source: Digital Trends, "Meta is changing how it judges employees after an unusual AI experiment," September 2026. This post walks the full arc, why the metric was doomed from the day it was written, and what an "impact" number would actually have to measure to survive contact with 85,000 engineers looking for the easiest way to move it. Seven months, one memo, and the leaderboard that broke it. The original policy, announced by Head of People Janelle Gale, framed the goal in aspirational terms: "As we move toward an AI-native future, we want to recognize people who are helping us get there faster." Source: Winbuzzer, February 2026. In practice, "helping us get there faster" collapsed into the one number that was easy to pull from a dashboard: how many tokens an engineer's AI usage burned. Meta even built supporting infrastructure for the mandate, a gamified learning platform called Level Up and an AI Performance Assistant pairing its internal Metamate bot with Google's Gemini, to help employees hit the new bar. None of that infrastructure measured whether the AI-assisted work was any good. It measured whether the AI was used. Employees found the gap immediately. An internal leaderboard, tracking token consumption across the company, started circulating; some reporting nicknamed it "Token Legend," other coverage of the same dashboard called it Claudeonomics. Source: Symplexia Labs, "Meta Pushes Its New AI Agent on Employees, but Eases Off on Tokenmaxxing," September 2026. Whatever its name, the mechanism was the same one this blog has already documented in enterprises that reward usage instead of outcomes: engineers re-ran full-context prompts instead of sending short follow-ups, kicked off agents to work multiple tasks in parallel, and generally optimized for the metric instead of the work, because the metric was the thing being graded. Thirty-day volume across the leaderboard hit 73.7 trillion tokens by April, up from 60.2 trillion. Source: Michael Parekh, "AI: Meta steps back from AI 'tokenmaxxing,'" AI-RTZ #1118, September 2026. The leaderboard's own creator took it down within days once it leaked publicly. Meta's CTO Andrew Bosworth tried to correct course from the top that same month, writing in an internal memo: "All motion is not progress; token usage alone is not impact." Source: Michael Parekh, AI-RTZ #1118, September 2026. The formal policy didn't move for another five months. It took until the week of September 2 for the performance-review language itself to catch up with what the CTO had already said in April: Meta confirmed it "will not use AI adoption dashboards or token counts to evaluate impact," and a company spokesperson told WIRED the labels were never really the point, contributions were. Source: Digital Trends, September 2026. Ten months, one mandate, one leaderboard, one reversal. Timeline: Meta announces AI-driven-impact performance reviews on February 4, 2026; a leaderboard tracking ~85,000 employees and 60 to 73.7 trillion tokens per 30 days leaks and is removed in April 2026; CTO Andrew Bosworth writes "token usage alone is not impact" the same month; the week of September 2, 2026, Meta formally confirms it will not use AI adoption dashboards or token counts to evaluate impact. Sources: Winbuzzer, Tom's Hardware, Michael Parekh (AI-RTZ #1118), The Information, Digital Trends. Why token volume was always the wrong proxy. None of this is really a Meta story. It's a Goodhart's Law story: any metric an organization rewards becomes a target the moment the people being measured can move it without moving the thing it was supposed to stand in for. Token count is exactly that kind of metric. It correlates weakly, sometimes negatively, with useful output, and it is trivially easy to inflate: run the same prompt twice, skip caching, kick off redundant agent branches, ask for a full-file rewrite where a two-line diff would do. This blog priced that exact failure mode in dollars already: agentic coding sessions can burn up to 1,000x the tokens of a single completion for the same underlying task, and a metric that rewards the bigger number rewards the worse engineering choice by construction. The mandate's authors weren't wrong that AI adoption was worth encouraging. They were measuring the thing that was easiest to instrument, not the thing they actually cared about. That's the same mistake enterprises make when they respond to a runaway bill with a flat per-seat spending cap: a cap changes how much volume gets through, not which model handled which request or whether the work got done. A volume floor and a volume ceiling are the same category of policy pointed in opposite directions, and both miss the number that was never on the dashboard: what did this request cost to produce a correct result, and would a cheaper path have gotten there anyway. What a real impact metric looks like instead. Meta's new language, results "can be supported by AI or other means," is deliberately unmeasured. That's the honest position if the alternative is measuring the wrong thing loudly. But an organization running thousands of engineers against agentic tooling doesn't have to give up on instrumentation entirely; it has to instrument the outcome, not the input. Two numbers survive the incentive problem the token-count mandate didn't: Cost per resolved unit of work (a merged PR, a closed ticket, a shipped answer), not cost per request and not tokens per request. This number falls when routing sends easy work to a cheap model and rises when someone burns a frontier model on a one-line fix, which is the opposite of what a raw token count rewards. First-pass correctness rate, whether the cheapest model that answered actually cleared the bar without a retry, a re-run, or an escalation. A verifier-gated cascade produces this number as a side effect of how it routes; a flat token count never could, because it has no concept of whether the tokens it counted were spent on a right answer or a wrong one that had to be redone. Both numbers are already sitting in a routing layer's request logs, because a router has to know the cost and the outcome of every call to make its next decision. The instrumentation Meta's memo implies engineering orgs now need isn't a new dashboard, it's the one a cost-aware router already produces: import openai client = openai.OpenAI( base_url="https://api.getnadir.com/v1", api_key="ndr_...", ) response = client.chat.completions.create( model="auto", # cheapest model that clears your quality floor, per request messages=[{"role": "user", "content": prompt}], ) Complete non-streaming responses report actual cost in nadir_metadata.cost.total_cost_usd. When a benchmark is configured and priced, nadir_metadata.benchmark_comparison.savings_usd reports the comparison. Neither number is "how many tokens did this engineer use" -- both are "what did this specific answer cost to produce," which is the number a performance review can actually defend. cost = response.model_extra["nadir_metadata"]["cost"]["total_cost_usd"] model_used = response.model An engineering org that logs cost-per-resolved-task and first-pass correctness by team doesn't need a leaderboard to know who's using AI well, and it can't be gamed by running the same prompt twice, because running it twice raises the cost of that resolved unit rather than the rank of the person who did it. The pattern is bigger than one company. Meta isn't alone in walking this back. Amazon shut down an internal leaderboard that ranked developers by token consumption in late May 2026, reportedly telling staff, "don't use AI just to use AI." This blog has already covered Microsoft, Uber, and Amazon hitting the same wall from the cost side rather than the incentive side, cancelling licenses and burning through annual budgets months early. What's different about Meta's reversal is that it's the first of the group to admit the volume metric itself, not just the resulting bill, was the design flaw. Gartner's own prediction that over 40% of agentic AI projects get canceled by 2027 names "unclear business value" as a leading cause right alongside cost. A company that spent ten months measuring the wrong number and still couldn't answer what its AI usage was worth is a live example of exactly that gap. What to check this week. Look for your own Claudeonomics. Any dashboard, leaderboard, or informal ranking inside your org that surfaces raw token or request volume by person or team is the same incentive Meta just spent seven months unwinding, whether or not it's tied to a formal review. Replace the volume number with a cost-per-outcome number. If your tooling can report tokens per engineer, it can report dollars per merged PR or dollars per resolved ticket instead. That number can't be inflated by running the same prompt twice; it gets worse when someone does. Track first-pass correctness, not just token spend. The 680x gap in per-employee AI spend this blog measured earlier is invisible on a volume chart and obvious the moment you ask which requests needed a second try. Don't swing to the opposite mandate. A flat per-seat cap is Claudeonomics in reverse: it still measures volume, just with a ceiling instead of a scoreboard. Neither one prices the request; only routing against a quality floor does that on every call. Meta spent ten months and one very public leaderboard finding out that "how much AI did you use" was never the same question as "was the work any good." The fix wasn't a better volume dashboard. It was giving up on volume as the signal and going looking for the number underneath it, cost against a resolved outcome, that a routing layer already has to compute just to route. Related reading Engineers defaulting to frontier models for every call, and what it costs at 5,000-engineer scale. Flat per-seat spending caps reduce volume. They don't change which model handled the request. Microsoft, Uber, and Amazon hit the same wall from the cost side, cancelling licenses and burning annual budgets early. A 680x gap in AI spend per employee is invisible on a volume chart. A verifier-gated cascade produces first-pass correctness as a side effect of how it routes. Sources: Winbuzzer, "Meta Ties Employee Performance Reviews To AI Usage Starting 2026," February 2026. Tom's Hardware, "AI cost crisis hits tech giants as employee tokenmaxxing backfires," May 2026. Michael Parekh, "AI: Meta steps back from AI 'tokenmaxxing,'" AI-RTZ #1118, September 2026. The Information, "Exclusive: Meta Tells Engineers AI Token Usage Won't Be Part Of Performance Reviews," September 2026. Digital Trends, "Meta is changing how it judges employees after an unusual AI experiment," September 2026. Symplexia Labs, "Meta Pushes Its New AI Agent on Employees, but Eases Off on Tokenmaxxing," September 2026.