I Thought I Saved $15,000 Using Claude.

Turns out I was asking the wrong question.

After checking my AI usage statistics, I did what every developer eventually does: I multiplied my token usage by the public API prices.

The result was ridiculous.

My usage would have cost roughly $15,000 through the API, while I had only paid a few hundred dollars for my subscription.

For a moment, I thought:

“I just saved fifteen grand.”

It’s a great headline.

It’s also misleading.

The hidden assumption is that I would have generated exactly the same number of tokens if every request had cost money.

I wouldn’t have.

Nobody would.

Flat-rate pricing doesn’t just make existing usage cheaper — it fundamentally changes how you work.

You ask bigger questions.

You let AI iterate.

You experiment.

You stop optimizing every prompt and start optimizing outcomes.

That’s why the “$15,000 saved” calculation tells only half the story.

The more interesting question isn’t what those tokens would have cost.

It’s whether those tokens would have existed at all under API pricing.

That realization sent me down a rabbit hole of AI pricing, token economics, subscription models, and why cheaper AI often leads to higher overall spending rather than lower bills.

I wrote the full analysis — including the calculations, the economics behind them, and why the commonly shared “look how much I saved” screenshots are more misleading than they appear.

You can read it here:

https://www.hoettler.com/en/blog/what-a-token-really-costs