There’s a new AI workplace trend getting attention: Tokenmaxxing.
The idea is simple: track how many AI tokens employees consume, sometimes through internal dashboards or leaderboards, and use that as a signal of AI adoption.
MBI Deep Dives argues this is not just a weird internal metric. More token usage means more real-world traces, which can become valuable training data for the next generation of models. ([<a href="https://www.mbi-deepdives.com/tokenmaxxing/" target="_blank" rel="noopener nofollow">MBI Deep Dives</a>])
That framing stuck with me because it changes the question.
This may not be only about developer productivity.
It may also be about data generation at organizational scale.
<h3>My take</h3>
Tokenmaxxing feels like the AI-era version of measuring developers by lines of code.
It can show activity.
It can show adoption.
It can show who is experimenting.
But it does not automatically show impact.
A developer burning 10M tokens might be:
<ul>
<li>building useful agent workflows</li>
<li>exploring model limits</li>
<li>generating traces that help the company learn</li>
<li>or just outsourcing confusion to a chatbot at high cost</li>
</ul>
The metric itself cannot tell the difference.
That’s the danger.
Once token usage becomes a leaderboard, people optimize for the leaderboard. That is classic Goodhart’s law: when a measure becomes a target, it stops being a good measure.
The more interesting metric is not:
<blockquote>
“Who used the most tokens?”
</blockquote>
It is:
<blockquote>
“Which AI-assisted workflows shipped better software, faster, with fewer regressions?”
</blockquote>
That means measuring outcomes:
<ul>
<li>cycle time</li>
<li>defect rate</li>
<li>review quality</li>
<li>incident reduction</li>
<li>customer impact</li>
<li>cost per useful task</li>
<li>reusable traces, evals, and agent workflows</li>
</ul>
Tokens are an input cost.
They are not the product.
<h3>The deeper question</h3>
There is a legitimate argument for encouraging heavy AI usage right now.
Early in a platform shift, waste can be part of learning. Teams need room to experiment, build agents, create workflows, and discover what actually works.
But if token usage becomes a status game, we may end up rewarding the wrong behavior:
<ul>
<li>more prompts instead of better prompts</li>
<li>more agent loops instead of better harnesses</li>
<li>more inference spend instead of better engineering judgment</li>
<li>more “AI activity” instead of better products</li>
</ul>
The best teams probably won’t be the ones that use the most tokens.
They’ll be the ones that turn tokens into repeatable leverage.

The official daily.dev AI community. Led by thought leaders and AI experts. Stay tuned for insights, news and discussions. 

Tokenmaxxing is an emerging workplace trend where organizations track AI token consumption as a proxy for AI adoption, sometimes via internal leaderboards. The author argues this mirrors the flawed practice of measuring developers by lines of code — it captures activity but not impact. Token usage can't distinguish between genuinely productive AI workflows and wasteful prompt spam. Goodhart's Law applies: once token counts become a target, people optimize for tokens rather than outcomes. Better metrics would focus on cycle time, defect rates, incident reduction, and reusable agent workflows. While heavy AI experimentation has value during a platform shift, the best teams will be those that convert token spend into repeatable, measurable leverage.

Tokenmaxxing: are we measuring AI productivity or just rewarding token burn?

Measuring AI adoption by token usage feels useful early, but dangerous once it becomes a target. It’s basically “lines of code” for the AI era. I’d rather see teams measure cycle time, defect rate, review quality, incident reduction, and cost per useful task. Otherwise we’re just gamifying inference spend.