Does token volume measure AI productivity or enable gaming?

Engineering-leadership newsletter writer Gregor Ojstersek reports that Meta built an internal leaderboard — bottom-up, "built by engineers and shared on the company's intranet" — that "tracks AI token usage of over 85k employees." The leaderboard awards tiered badges "from bronze, silver, gold, platinum, to emerald," and treats the top 250 ranked employees as "power users" eligible for titles like "Session Immortal" and "Token Legend." Ojstersek names the resulting behavior "tokenmaxxing": treating "token consumption as a benchmark for productivity and a competitive metric to determine if an employee is 'AI-native.'"

Ojstersek's stated worry is a measurement-substitution problem: "if token usage becomes the metric, then token usage becomes the goal." He reports that once volume itself carries status, employees "let AI agents run continuously for hours to perform research tasks, maximizing token consumption," and that the easiest way to "rack up tokens" is "keeping a chat context going for a long time," feeding it "multiple repos for extra points," and pasting in as much text as possible — none of which requires the work to be useful. The excerpt also reports a set of executive statements that reinforce the same substitution, though it names a speaker only once: Ghodsi, credited with treating an engineer's $7,000 in two weeks of token spend as something to "applaud" rather than question. Other, unattributed lines in the same excerpt claim a top engineer who spent "an amount equivalent to their salary" on tokens saw "a productivity increase of up to 10 times," with "no upper limit" and the stated goal being to "maximize your token throughput." Ojstersek presents these statements as a driver of the leaderboard culture, not as independent evidence that more tokens produced more value.

This cuts against Are multi-agent systems actually intelligent coordination or just token spending?, which treats heavy token consumption as mostly wasted compute rather than a sign of capability; Meta's leaderboard inverts that reading by treating the same volume as a status marker regardless of what the tokens accomplished. It is also a cruder instrument than How much progress have AI agents actually made on NanoGPT?, which only credits an agent for dollar-for-dollar improvement matched against a human doing the same task on the same budget — exactly the discipline a raw-volume leaderboard skips.

The excerpt gives no outcome data tying leaderboard rank, badge tier, or the cited dollar figures to any actual output quality, business result, or measured productivity; the $7,000, "10x," and "equivalent to salary" figures are anecdotes relayed secondhand ("some unofficial data I have seen based on my research," "I am also hearing the following"), not controlled comparisons. Ojstersek's own framing — "stupid and easily gamed," comparable to "measuring lines of code or using story points" — is a judgment, not a finding, though the specific gaming behavior he describes (idle agents run for hours, context padded with unneeded repos) is a concrete, checkable claim about how people respond once a proxy metric carries consequences, up to and including termination at the unnamed company he says he works for.