The perils of tokenmaxxing: How to govern AI spend without sacrificing speed

I write about tech and AI for a living, but nothing has made me yearn for the Butlerian Jihad more than learning (against my will) about the term "tokenmaxxing." And if I have to know what that means, you do, too.

In early 2026, companies started publishing internal leaderboards ranking employees by how many AI tokens they consumed. At Meta, the board was called "Claudeonomics," and it handed out digital badges with extremely not-dorky titles like "Cache Wizard" and "Model Connoisseur." The highest-ranked individual user averaged 281 billion tokens—with costs running into the hundreds of thousands of dollars—in a single period. Amazon had a similar leaderboard, which it eventually shut down in May 2026 with an internal correction: "Don't use AI just for the sake of using AI." What a concept.

Here's everything you need to know about tokenmaxxing, why companies are pivoting hard away from it, and how to set AI usage goals that actually work for you (and not just against your town's electrical grid).

Table of contents: 

• What is tokenmaxxing?

• The problem with tokenmaxxing

• Tokenmaxxing vs. tokenminning: The overcorrection

• What should you actually measure instead?

• How to stop paying inference prices for deterministic work

What is tokenmaxxing?

When you type a prompt into an AI model, the model doesn't read your words the way you do. It breaks your text into small chunks called tokens, which are roughly three-quarters of a word on average. So the word "tokenmaxxing" takes two tokens, a short paragraph is around 75 to 100 tokens, and a long document can run into the thousands. Every prompt you send and every response the model generates gets processed as a stream of these chunks. Most AI providers bill by how many tokens flow through their systems, either through monthly subscription caps or per-token API pricing, so the more your team uses AI, the more tokens burn—and the bigger the line item gets.

The term "tokenmaxxing" borrows the (highly questionable) "-maxxing" suffix that's been floating around in the darker corners of the internet for years. My poor terminally online eyes have been subjected to everything from looksmaxxing to sleepmaxxing, gymmaxxing, and even fibermaxxing (don't ask). The point seems to be to pick one metric, push it as hard as possible, and treat the number itself as the point, whether or not anything useful comes out the other side. 

Applied to AI, that translates to a simple and extremely gameable idea: the more tokens your team burns, the more they're "using AI." And the more they're using AI, the better they're doing. Obviously.

That logic had a certain appeal (to some) in 2025 and early 2026, when AI adoption was being tracked, measured, and often, explicitly mandated. Shopify CEO Tobi Lütke sent a memo to employees stating that anyone who wasn't using AI in their work could expect to have a conversation with their manager. Nvidia CEO Jensen Huang said publicly that if an engineer with a $500,000 salary didn't consume at least $250,000 worth of tokens in a year, he'd be "deeply alarmed." What's deeply alarming is in the eye of the beholder, I guess. 

So companies measured what they could measure. Since tokens were the only input every AI provider could meter consistently across models and platforms, token spend became a convenient proxy for AI engagement—and, by extension, for innovation. If you were burning tokens, you were in the game. Leaderboards went up, and people found creative ways to rank higher. 

But that all changed when the bills started rolling in.

The problem with tokenmaxxing

People found two ways to tokenmaxx (wow, it sounds even worse as an infinitive). The first is what you might call usage theater. The name of the game was using as much AI as possible at every turn—which meant using longer prompts and leaving AI agents running in the background (even when they weren't doing anything). The output doesn't necessarily improve, but the number of used tokens does get higher.  

The second strategy is model defaulting. When the organization is pressuring you to use AI as much as possible, teams naturally reach for the most powerful models available. The problem is that the most powerful models are also the most expensive, and a lot of the work getting routed through them doesn't need that kind of horsepower. Summarizing a structured weekly report and synthesizing six months of unstructured customer interviews aren't the same kind of task, but if your default is always the most powerful Claude or OpenAI model, you're paying frontier prices across the board. 

Token volume is a vanity metric. It's a straightforward case of measuring the wrong thing. The closest analogy is measuring developer productivity by lines of code, which is a benchmark the software industry spent decades learning to reject because it rewards padding over efficiency. Just like you can write 500 lines of code that do the work of 50, you can burn 500,000 tokens and produce nothing worth keeping. Volume and value aren't the same thing, and when you optimize for volume, that's usually all you get. (I learned this lesson years ago in college when I bought a bulk pack of off-brand instant coffee because it was technically more coffee. But hey, everyone learns at their own pace.)

There's also the longer-term math problem as the inevitability of rising costs. Both Anthropic and OpenAI are burning through investor money and heading toward IPOs, which means the below-cost pricing era is quickly coming to an end—and wasting tokens on inefficient work will become even more painful.

We're coming for your job, AI

Listen to enough of the AI discourse, and you might pick up on an assumption that human judgment isn't an asset but a bottleneck to be engineered around. But if tokenmaxxing has taught us anything, it's that the common refrain about how AI will "replace all jobs" is a little too optimistic (if wishing for a dystopian nightmare is optimism). 

In fact, using AI as indiscriminately as tokenmaxxers do might mean paying more for worse results than you'd get from hiring a human to do the same work. AI works best when a real person is making deliberate choices about where to deploy it, what model to use, and what a good output actually looks like. The human brain is still often the most cost-efficient processor in the room—so don't count us out yet.     

Tokenmaxxing vs. tokenminning: the overcorrection

Many organizations' gut reaction to tokenmaxxing chaos is to simply limit subscriptions, add usage caps, or pull back on tools. This trend has an equally ridiculous name: tokenminning. 

I'm not saying it's not understandable. I myself often think longingly of the days when AI was just a twinkle in Isaac Asimov's eye. 

But the problem is that blanket cuts don't distinguish between workflows where AI earns its cost and those where it doesn't. They just penalize everything equally, including the work that actually justifies the cost. If someone on your team is using AI to synthesize a hundred unstructured customer interviews into actionable themes, or to draft first-pass responses to complicated support escalations that a rep then refines and sends, cutting their access because other people were running leaderboard games isn't really solving the problem. You'll just be trading one overreaction for another.

What should you actually measure instead?

The solution isn't complicated, but it does require a mindset shift: instead of measuring AI input, start measuring its output. 

Here are a few examples of AI ROI metrics you could track instead. That way, you can save your leaderboard-building skills for gamifying actually fun things, like tracking how many times people say "optimize" in a team meeting. 

Time saved on high-judgment work

When used correctly, AI can earn its cost by saving your team meaningful time on tasks that used to require manual analysis, like research synthesis, first-draft writing, and data interpretation. But if your team is spending more time reviewing and fixing AI output than they would have spent just doing the work themselves, that's a sign something's miscalibrated.

Decisions made with fewer errors, not just faster

I can do all of our dishes in about five minutes. But that's because I cram everything into the dishwasher without pre-rinsing and hope for the best. My husband might take three times as long as I do to clean the kitchen, but everything comes out sparkling clean the first time. Speed isn't always the best metric. 

AI that helps someone make a bad decision faster isn't an improvement. A token that produces output a human immediately throws away is a wasted token by definition—and if that's happening consistently, it usually means the AI is being used in the wrong part of the workflow. 

The goal is to improve operational efficiency with AI. So instead of measuring how fast you can get to an output, you'll need to be more holistic about how much more efficient your team is overall with AI's help. 

Customer-facing benchmarks

Maybe you use AI-powered marketing personalization or product recommendation tools. Or, maybe AI is helping you brainstorm more business opportunities or enabling faster product development cycles. Is AI measurably helping you increase conversion rates, leads, or traffic? 

Instead of tracking new AI-specific (and extremely abstract) metrics like tokens, keep tracking the metrics that have always mattered for business growth. If you're seeing better numbers in those areas after AI implementation, you're doing something right.  

How to stop paying inference prices for deterministic work

If your default is to run everything through AI because you have an AI subscription and you're supposed to be using it, you end up paying inference prices for logic that a deterministic automation would do better and for free. A big chunk of unnecessary AI spend comes from using an AI model to do work that doesn't actually require inference, like:

• Sending a confirmation email when a transaction completes

• Flagging an expense report that exceeds a threshold

• Routing an incoming lead to the right sales rep based on company size and territory

These are all deterministic workflows with exactly one right answer, and a model doesn't need to "think" to get there. A rule can handle it instantly and get it right every time. 

Non-deterministic work is different: it involves ambiguity, context, or judgment that a rule can't anticipate. For example, an AI model can summarize a messy batch of open-ended survey responses or write a personalized reply to a customer escalation where tone and nuance matter.  

Zapier is built for exactly this split. You can combine deterministic automation steps with AI by Zapier in the same workflow, routing each step to whatever actually handles it best. If you're using the default AI model, an AI step costs the same as any other traditional automation step. If you need more AI power, you can tier up, and you can swap models whenever a better price option comes along—without rebuilding anything. 

Tokenmaxxing is out. Valuemaxxing is in. (Sorry)

AI tokens were visible and countable, so they became the metric—and once something becomes a metric, people will find a way to game it. They always do. But then the bills arrived—and everyone discovered, at significant expense, that volume and value aren't even close to the same thing. You can burn a quarter million dollars in tokens and have nothing to show for it but a leaderboard badge that says "Cache Wizard."

The irony is that the organizations spending the most on AI often got the least out of it, because they optimized for activity instead of outcomes. But more and more companies are learning that the best AI use is unglamorous, useful, and often invisible. Successful AI adoption simply helps you spend less time on the work that doesn't need you, and more time on the work that does. And that gives me hope.

I'm not quite ready to call for the Butlerian Jihad. But if you're still running leaderboards, I'm at least thinking about it. 

Related reading: 

• AI in business: 38 statistics + insights & use cases 

• How to measure AI adoption: Key metrics to track

• AI at Zapier: How we use artificial intelligence to streamline work

• Tool sprawl limits AI integration for 70% of enterprises   

• AI workflows: How to actually use AI in your business