I Asked Codex to Cancel a Subscription—and It Even Got Me a Refund

To set the scene and add a bit of storytelling: I had just got back from the pool, feeling pretty relaxed, when I found an email telling me I'd been charged $50 for the annual renewal of my Medium subscription. I've been subscribed to the platform for years, but last year I realised I practically NEVER read the articles anymore. So, as usual, I decided I would cancel—and then probably went off to do something else without ever acting on the thought.

The $50 Medium renewal charge that triggered the cancellation request.

I have to admit I have a soft spot for the platform. In the past, I considered it one of the most reliable places to find tutorials and genuinely useful information, alongside Stack Overflow. But over the last few years it has stopped being as useful to me, and I feel the quality of the articles recommended in its newsletter has fallen dramatically.

That said, charging you without warning strikes me as a bad practice. Could they remind you a few days before the subscription renews, in case you want to cancel? Of course. Would a lot of people cancel? Absolutely—and that is why they do not.

So I decided I wasn't just going to cancel the subscription; I was also going to ask for a refund. I was feeling pretty bold, mainly because I already had a plan: I wasn't going to do it myself. AI was.

Anyone who has tried this knows it is one of the highest-friction processes imaginable. When you want a user to complete an action, you reduce the number of clicks it takes. Here, they make you click through six or seven buttons for every tiny bit of progress because, obviously, they know you will give up, abandon it, forget about it… until next year, when the next bill arrives.

But this time they had run into someone paying for the most expensive Anthropic and OpenAI subscriptions (plus a few others), constantly looking for ways to spend tokens. A light bulb went on.

The short prompt that asked Codex to handle the cancellation and refund.

One screenshot and a prompt just over two lines long. A state-of-the-art frontier model. Full access to my computer—use whatever you need. I trust you, OpenAI. Get back what is mine; I know you will not let me down.

I pressed Enter, went to lunch, and then took a proper summer nap.

It was not exactly epic. After just 1m 54s, the process stopped because it needed me to sign in to my Medium account, so it left the login screen open and waited. I signed in, told it to continue, and at 2m 49s it asked me to confirm my PayPal account. I did, and only 26 seconds later it asked me to click “Confirm cancellation”.

Finally, 1m 8s after I confirmed the cancellation, it sent me this message:

Codex confirms that the subscription was cancelled and the refund requested.

In total, it took about five minutes and it was telling me the subscription had been cancelled and the refund requested. The speed honestly surprised me. Had I done it manually, I do not think it would have taken me less than 30 minutes—and probably longer.

And a warning for every subscription business: my agent has learned that cancelling and requesting a refund are two separate things. It may sound like a joke, but persistent memory and automated learning are essential for efficient AI-driven processes. That means next time it can do the job in three minutes instead of five, and it already knows which PayPal account is mine, so I will not need to tell it.

At that point, all that remained was to wait and see whether OpenAI would actually deliver what it had promised. And it absolutely did. This is the email I found first thing the following morning:

The following morning, PayPal confirmed the $50 refund.

I find it incredible. I have to admit that at first I intended to try it without any real expectations. In fact, I had probably already resigned myself to it not working. But it genuinely surprised me because this is a fairly common task, it solves a problem we ALL have, and I think the same approach could be extended to other use cases with only small changes.