Inside Meta’s token culture: 'Claudeonomics'

Meta employees briefly had an internal scoreboard that ranked the company’s top 250 artificial intelligence users by token consumption, and the board was nicknamed “Claudeonomics” after Anthropic’s Claude model. The Information reported that Meta has now taken the board down after the usage data started circulating outside the company. (theinformation.com) A token is the tiny unit an artificial intelligence model counts when it reads a prompt or writes a reply, and OpenAI says one token is roughly four characters of text. Anthropic’s own developer docs say token counts are estimates, which means a leaderboard built on tokens is measuring volume, not cleanly measuring useful work. (help.openai.com) (platform.claude.com) The striking number in the leak was scale: more than 60 trillion tokens used across Meta in a 30-day span. The same reporting said the company has more than 85,000 employees in the system and that the single top user logged about 281 billion tokens in that month. (theinformation.com) (the-decoder.com) The board did not just list names. It handed out status labels including “Token Legend,” “Session Immortal,” “Model Connoisseur,” and “Cache Wizard,” which turned routine model usage into something closer to a company fantasy league. (aol.com) (theinformation.com) That setup fits a wider Silicon Valley habit that employees have started calling “tokenmaxxing,” where using more artificial intelligence is treated as proof that you are moving faster. Several reports said some workers were even leaving agents running to keep their numbers high, which shows how quickly a cost meter can become a status game. (the-decoder.com) (blog.tmcnet.com) The name “Claudeonomics” also says something about Meta’s tool stack. Even though Meta builds its own Llama models, the nickname points to heavy internal use of Anthropic systems for coding and knowledge work, which matches the broader industry pattern of companies mixing in-house models with outside ones that perform better on specific tasks. (theinformation.com) (anthropic.com) The hidden argument inside the leaderboard was that more prompts produce more output, the same way more compute can produce a stronger model. That can be true in a rough sense, but token counts are a noisy proxy because a short, well-aimed prompt can solve a bug faster than a million-token agent left spinning overnight. (platform.claude.com) (help.openai.com) Meta’s decision to remove the board suggests the company liked aggressive experimentation more than it liked public visibility into the bill. Once employees outside the top ranks could see the scale, and once outsiders could too, a playful internal dashboard started to look like a window into how expensive the artificial intelligence arms race has become inside one of the world’s biggest tech companies. (theinformation.com) (techmeme.com)