Tokenmaxxing: The Dangerous New AI Trend Businesses Need To Stop

Some companies are encouraging employees to burn through as much AI as possible, and even ranking them by how many AI tokens they consume.

Welcome to the strange world of “tokenmaxxing.”

Tokens are essentially the currency of generative AI. They are the small chunks of text that AI models process when you enter a prompt or receive an answer, and companies such as OpenAI, Anthropic and Google typically charge businesses according to how many they use.

That makes token consumption useful for tracking AI costs. The trouble starts when businesses treat it as a measure of AI adoption, productivity or enthusiasm.

This has led to employees deliberately maximizing their token usage to climb leaderboards, hit targets or prove that they are embracing AI.

And as businesses pour billions into AI, that seemingly harmless metric could become a surprisingly expensive mistake.

So What Is It?

Although its exact origin is unclear, the first time many people heard the term was likely to have been in reporting of Meta’s Claudeonomics leaderboard, built by an employee to rank staff by their token usage.

With workers facing pressure from above to show they were using AI to drive productivity, or at the very least just using AI, the leaderboard showed the top 250 token users. For a short while, at least, until it became public knowledge and was quickly deleted.

This could be seen as a private joke among colleagues. But it was reflective of a culture where the easily captured metric of token use was being leaned on too heavily. Particularly when it wasn’t backed with evidence that it was linked to productivity, customer experience improvement or revenue growth.

Meta was by no means the only tech giant affected. Reports emerged that workers at Amazon had worked out how to game their own leaderboard in order to appear keener adopters of AI than they actually were.

Why It’s A Bad Idea

Today, businesses are keen to show that they’re using AI, and counting tokens seems like the easiest way to do that.