Meta employees are engaged in a crazy competition: who uses the most tokens?

Within Meta, burning through the most AI compute power is becoming a new status symbol.

On April 6, The Information reported that an AI usage leaderboard named “Claudeonomics” appeared within Meta Platforms. Built spontaneously by employees on the company intranet, the leaderboard tracks AI token consumption for more than 85k employees and lists the top 250 “super users.” Employees near the top can earn titles such as “Session Immortal” and even “Token Legend.”

The leaderboard name “Claudeonomics” comes from Claude, Anthropic’s flagship product, an AI startup. A copy of the leaderboard obtained by the media shows that in the past 30 days, the total token usage recorded by the leaderboard exceeded 600 trillion. The highest-ranking individual users averaged 281 billion tokens consumed each—depending on the model type, the cost of this usage could be as high as several million dollars.

According to Anthropic’s latest publicly available pricing, the average price for input and output tokens for its Claude Opus 4.6 model is about $15 per million tokens. Using this estimate, 600 trillion tokens would correspond to a cost of about $900 million. However, it is not clear which models Meta actually uses and at what prices it purchases them.

“Burning Tokens” becomes a new yardstick for productivity

This phenomenon reflects the “tokenmaxxing” culture taking off in Silicon Valley—treating token consumption as a benchmark for measuring productivity, and as a competitive metric for judging whether employees are “AI-native.”

Tech executives are endorsing it.

NVIDIA CEO Jensen Huang said last month that if an engineer earning $500k a year spends less than $85k annually on AI tokens, he would be “deeply concerned.”

Meta CTO Andrew Bosworth said at a technology conference in February this year, according to Forbes, that a top engineer would spend an amount equal to their salary on AI tokens, with productivity gains up to as high as 10x. Bosworth said bluntly: “It’s a buy that pays off with no risk—keep doing it, there’s no ceiling.”

Last month, former Tesla and OpenAI top AI scientist, and current head of an AI education startup, Andrej Karpathy, also said on a podcast: “The key is token. What’s your token throughput? How much token throughput can you mobilize?”

How the leaderboard works

Employees can track their own consumption on the leaderboard, compare horizontally with colleagues, and receive gamified rewards—from bronze, silver, gold, platinum to emerald badges—along with achievement titles such as “Model Connoisseur” and “Cache Wizard.”

According to two current employees, some employees keep their AI agents running for hours to carry out research tasks in order to maximize token consumption and climb the rankings.

Meta official also has an independent token usage dashboard for software engineers; employees in other roles can also view their usage. Notably, according to a source familiar with the matter, neither Zuckerberg nor Bosworth himself made the top 250 super user leaderboard.

At the tooling layer, in addition to using Anthropic, OpenAI, and Google models, Meta employees can also use internal development tools, including Meta’s version of OpenClaw (internally called MyClaw), as well as Manus, which Meta recently acquired.

A Meta spokesperson said: “As everyone knows, this is the company’s priority—we focus on using AI to help employees get their daily work done.”

Voices of doubt: does consumption equal productivity?

This competition is not without controversy.

Joe Weisenthal, a media personality at Bloomberg, directly asked on X: “What does it even mean to measure productivity by total token consumption?”

He went on to mock it further, saying: ‘It really has that “real backyard steel furnaces vibe”’—meaning the frenzy of chasing numerical metrics while ignoring real quality, which is very much like wasteful resource usage with no regard for cost.

This skepticism points to a core issue: token consumption is an input metric, not an output metric. Just as measuring employee work efficiency by the number of sheets of paper printed doesn’t mean more tokens burned leads to more results. The behavior of some employees letting AI agents “run idle” for hours to climb the leaderboard is precisely evidence that this metric has room to be “gamed” with “data brushing.”

In response, well-known tech analyst Noah Brier offered a different view: “I don’t think this is justified, but when you try to turn a giant organization like Meta, sometimes you have to make a ‘purposely overcorrect’ move.”

However, Weisenthal immediately followed up, asking: “Even so, what exactly are they trying to turn around—the employees’ work habits, or the company’s money-making model?”

Still, from a market perspective, the phenomenon itself sends a clear signal: enterprise AI consumption is expanding at a pace far faster than expected. Even for just Meta alone, monthly estimates of AI compute spending could be close to the same order of magnitude of $900 million. For cloud computing and AI infrastructure suppliers, that means sustained demand growth.

Risk disclosure and disclaimer

        There are risks in the market; invest cautiously. This article does not constitute personal investment advice, and it does not take into account any specific investment objectives, financial conditions, or needs of individual users. Users should consider whether any opinions, viewpoints, or conclusions in this article align with their specific circumstances. If you invest based on this, you bear responsibility.