The Tokenmaxxing Hangover: When AI Becomes More Expensive Than People

AI isn't getting worse. It's getting expensive.

For the past two years, the AI industry had one simple message:

Use more AI.

More prompts. More agents. More reasoning. More tokens.

Some companies even started measuring employee productivity by the number of AI tokens they consumed. It became known as tokenmaxxing, a term that was likely first popularized by technology journalist Mike Pearl (Gizmodo) in early 2026 when he described organizations evaluating engineers by how fast they burned through LLM tokens.

Today, the mood is changing. The invoices are arriving. And suddenly everyone remembers that AI was never free.

From experimentation to expensive reality

The recent Belgian newspaper article in De Standaard caught my attention.

Companies are seeing their AI bills increase dramatically. Some developers are consuming thousands of euros worth of AI tokens every month. Organizations that enthusiastically encouraged AI experimentation only months ago are now introducing budgets, limits and governance.

Some are even asking an uncomfortable question:

Would a human actually be cheaper for this task?

Not because AI doesn't work. Because using the most powerful model for every task isn't always the most economical solution.

The Meta leaderboard

Perhaps the most striking example came from Meta.

According to several reports, employees created an internal leaderboard—nicknamed "Claudenomics"—ranking engineers by how many AI tokens they consumed. The assumption was simple: more tokens meant more AI adoption, which meant more productivity.

The numbers quickly became staggering. Total usage: 60.2 trillion tokens in 30 days.

One OpenAI engineer reportedly processed 210 billion tokens in a single week. Other companies reported individual monthly token bills well into six figures. One Anthropic customer allegedly generated a Claude Code invoice of around $150,000 in a single month.

Those stories generated headlines.

They also generated something else: questions.

Because token consumption is easy to measure. Business value is much harder.

When a KPI becomes the goal

This happens more often than people realize.

A metric that starts as a useful signal slowly becomes the objective itself.

Instead of asking whether AI actually improved the business, organizations started measuring how much AI people consumed.

The questions shifted from:

• Did we ship software faster?
• Did we improve customer satisfaction?
• Did we reduce operational costs?
• Did employees spend less time on repetitive work?

to a much simpler question:

How many tokens did you consume?

That's like measuring a delivery company's success by how much diesel its trucks burn. Or evaluating a factory by its electricity bill.

Consumption is not value.

The first AI scepticism isn't about AI

What's happening today isn't really an AI backlash.

It's a reality check.

For nearly two years, AI providers encouraged unlimited experimentation. Startup credits, generous subscriptions and investor-funded growth created the impression that AI capacity was almost infinite.

Now the economics are becoming visible.

Premium reasoning models are expensive. Agentic workflows multiply token consumption. Long-running coding agents can generate enormous compute bills.

As organizations begin tracking these costs, the conversation is naturally shifting from "How much AI are we using?" to "What value are we actually creating?"

That's a much healthier discussion.

Even Nature says: stop tokenmaxxing

The discussion has become serious enough that Nature Machine Intelligence recently published an editorial with a clear message:

Stop tokenmaxxing and deploy AI sensibly instead.

The editorial argues that organizations are caught in a race to adopt agentic AI without always considering the financial, environmental and human consequences. More AI usage is not automatically better AI usage.

Instead, success should be measured by real outcomes while maintaining human oversight and careful evaluation of AI-generated results.

That feels like common sense.

Unfortunately, common sense often arrives after the invoice.

The conversation is finally shifting

The first wave of enterprise AI was all about adoption. Companies wanted employees to experiment. The more AI you used, the more innovative you were considered.

Today, a different conversation is emerging.

Finance departments are asking about ROI. IT departments are asking about governance. Business leaders are asking whether the additional AI cost actually creates additional business value.

That's a sign the market is maturing. The goal is no longer to use as much AI as possible. The goal is to use the right AI, where it creates measurable value.

Why we built ReplyFabric differently

This is one of the principles we had from day one. Our customers don't pay per token. They pay a fixed monthly subscription. That means the optimization challenge is ours, not theirs.

Every unnecessary token we consume is our cost, so we've spent a lot of engineering effort optimizing every step in the pipeline. Not every task deserves the most capable reasoning model. Not every prompt requires thousands of tokens.

As we often say internally:

You don't need Claude Mythos for language detection.

Or, to put it another way:

You don't drive a Ferrari to pick up bread from the bakery around the corner.

A lightweight model can detect the language of an email almost instantly. Another model can classify the intent. A different model can extract structured information. Only when an email genuinely requires deeper reasoning do we call a larger model.

Choosing the right model, at the right moment, for the right task allows us to deliver high-quality results while keeping AI efficient and predictable.

Our customers don't have to worry about exploding token bills.

That's our job.

The next generation of AI isn't about more tokens

I don't believe companies should become more sceptical about AI. They should become more disciplined. The first generation of enterprise AI optimized for adoption. The next generation will optimize for value.

Instead of asking:

• How many tokens did we consume?
• Which model did we use?
• How many AI agents are running?

Companies will increasingly ask:

• How much time did we save?
• How much faster do customers receive an answer?
• How much manual work disappeared?
• What's the return on every euro invested?

Those are business metrics. Everything else is just implementation.

I suspect that, a year from now, we'll hear much less about tokenmaxxing and much more about value.

And that's exactly where enterprise AI should be heading.

---

Sources

For readers who want to dive deeper into the discussion:

• Download Forbes article (pdf): Is The Cult Of 'Tokenmaxxing' Just Another Fad Or The New Normal?
• Download Nature article (pdf): Stop 'tokenmaxxing' and deploy AI sensibly instead