The Sunday I left AI sorting my photos and it cost me almost €1,000

It was Sunday. My photo library had been a mess for years, and I had a simple plan: let AI sort it out while I spent the day out. By type, by person, by whatever made sense.

I chose ChatGPT Astra, which had just come out. It was the biggest model at the time. If it was going to work on its own, I might as well pick the smartest one. I gave it the task, closed my computer and left.

On Monday, I opened my work email. Three notifications in a row. Three automatic API credit top-ups of €300 each, one after another. €900 in a single Sunday.

What had happened

Astra hadn't been idle. It had done exactly what a large model does when nobody sets any limits: think big.

It designed a five-step system. At each step, it saved a complete copy of the library, just in case. Five steps, five libraries. Around 80 GB of disk space taken up by copies of the same photos.

And every step cost tokens. Lots of them. Read, classify, check, check again. All with Astra, the most expensive model, all day long and with nobody watching. Whenever the balance ran out, the automatic top-up added another €300. Three times.

There was no technical error. The system worked exactly as it had been configured to. The mistake was mine.

The three mistakes

One: using a sledgehammer to crack a nut. Sorting photos doesn't need the most powerful model on the market. It needs one that follows clear instructions without dreaming up an engineering project. A large model doesn't do a simple task better. It makes it more expensive, and sometimes more complicated.

Two: leaving it on its own with no brakes. Back then, I didn't have the controls on my phone that I have now: usage alerts, automatic stops, a notification when something spikes. I gave it complete freedom and a whole Sunday. It did what anyone would do with complete freedom and a whole Sunday.

Three: automatic top-ups. Paying for AI as you use it, through the API, is not the same as paying for a monthly subscription. With a subscription, there's a ceiling on spending. With the API and automatic top-ups enabled, your card sets the ceiling. I didn't know that. I learned it on Monday, reading my email. It cost me €900.

How I did it the second time

I repeated the same task with a different approach. A large model, Opus, just to direct the work: understand what I wanted, draw up a short plan and review the result. And ChatGPT Terra, a medium-sized and much cheaper model, for the mechanical work of moving and classifying.

It was tens of times faster and cheaper. And without five copies of anything.

The difference wasn't in the AI. It was in how I used it.

What to do before leaving AI to work on its own

• Choose the model for the task, not just in case. The large one for thinking, planning and reviewing. The medium-sized or small one for doing the work. And if a simple program can solve it without AI, better still.
• Set a spending limit before you leave. Turn off automatic top-ups or put a cap on them. A balance that runs out is a warning. A balance that tops itself up is a bill.
• Enable alerts and automatic stops. A notification on your phone when usage goes above one threshold. A stop if it goes above another.
• Tell it what it mustn't do. “Don't duplicate files.” “Don't save intermediate copies.” “Stop and ask me if the plan has more than two steps.”
• Check on it at the start. Ten minutes watching what it does in the first half hour can save you a whole Sunday of surprises.

The lesson that stayed with me

AI doesn't get tired, doesn't get bored and doesn't look at the bill. Those three things are down to whoever puts it to work.

Since that Sunday, before leaving a bot to work on its own, I ask myself one question: if this goes wrong while I'm not watching, how much could it cost me? If I don't know the answer, I don't leave.