Tokenmaxxing - AI use as status symbol

Companies using tokens as a metric for productivity risk falling foul of Goodhart’s Law

In some businesses the consumption of AI tokens is used as a proxy for adoption. Competition to maximise the consumption of AI (“tokenmaxxing”) may have implications for budgets, governance and productivity.

Tokens are small strings of text (around four characters) that form the basic units of an LLM’s inputs and outputs. Tokenmaxxing describes the attempt by users to consume as many tokens as possible in order to demonstrate their AI usage. Some people reportedly even run parallel agents in the background to maximise their token throughput.

Why? Well, tokens make AI usage visible; visibility creates comparability, and at a time when most businesses are encouraging employees to use AI, coming top could mean rewards.

However, using token count as a core measure of productivity is likely to lead to negative outcomes, not least high costs. AI companies are starting to bill by the token and reports suggest that monthly bills for AI usage in some tech companies are already reaching six-figure sums.

Goodhart’s Law in practice

Goodhart’s Law states: “When a measure becomes a target, it ceases to be a good measure.” The following examples show that using token consumption as a target can lead to pointless consumption.

Employees at Meta Platforms were engaged in an intense internal competition to maximise AI usage, vying for a place at the top of the leaderboard with titles such as “Token Legend”. More than 60 trillion tokens were recorded within 30 days. The tool was shut down within 48 hours in early April 2026 after data from the internal leaderboard was leaked.

An example from the payments sector illustrates that the issue is not confined to Silicon Valley: Business Insider reports that Visa consumed around 1.9 trillion tokens per month in March - roughly double the figure for February - and was recognising teams for achieving faster results through the use of AI.

The situation becomes particularly sensitive where token or tool usage is not only aggregated but also visible on a personal basis.

JPMorgan is said to be using internal dashboards to track the use of tools such as GitHub Copilot and Anthropic models and to categorise user groups.

Disney is also reported to have an “AI Adoption Dashboard” that reports token and request figures for tools such as Cursor and Claude - complete with a leaderboard. One user is reported to have prompted Claude around 460,000 times over a period of nine working days; in reality, only autonomous agents can generate such invocation rates.

Outcomes before activity

People will always game a system to achieve rewards. This means that tokenmaxxing is unlikely to lead to the desired outcome. Particularly in AI coding, more generated code can also mean more rework.

Blogger Gergely Orosz notes: “Tokenmaxxing: great for AI vendors, bad for everyone else.”

But as usual, it’s not just black and white. Tokens are useful as a measure of cost and capacity. But as a status symbol or leaderboard currency, they risk distorting productivity.

Anthony Moisant, CIO of Indeed, emphasised in an interview that whilst his company tracks token usage in the background, it is certainly not interested in leaderboards. “I think anytime we have a metric or a measure like this that is part of an incentive system, we create that perverse incentive and people start to do things," Moisant told Business Insider.

Moisant prefers to focus more on outcome-oriented metrics (e.g. speed of shipping, customer impact), rather than measures of activity.

Mikhail Parakhin, CTO at Shopify, explained in detail in an episode of the Latent Space podcast what it actually looks like when a 20-year-old software company with an unlimited Opus 4.6 token budget goes all-in on AI. He said he has noticed the top developers using ever more tokens than the rest of the cohort.

“It feels not ideal, to be honest, but maybe it's okay. We'll see. Because take it to the limit. … if this rate of separation continues, there will be one person consuming all the tokens.”

This article first appeared on Computing Deutschland.