Cursor and the danger of overpromising

A couple weeks ago, I temporarily switched to using GPT-5 as my exclusive coding assistant model, via Cursor. This was prompted by two developments:

First of all, I was curious to try GPT-5, which had some buzz as a new contender the latest state of the art coding model. But I haven't wanted to go through openAI's API license verification process, because I really don't like an AI startup asking to see my driver's license. Using it through my Cursor subscription seems to be an end-run around that, for now?

But that only explains the specific choice of model. If GPT-5 hadn't recently been released I probably would have done the same with my daily go-to model (Sonnet 4), or tried out Gemini some more, or whatever.

The urgency came from this: My $20 pro plan Cursor usage was about to reset, and I hadn't run out of monthly credits yet. And the Cursor user forums were exploding with people outraged that their previous normal usage was suddenly hitting new limits.

So I really wanted to get some clarity about exactly what "at least $20 of included usage per month" might currently mean to Cursor.

So: I took the opportunity to try to make some headway on some thorny parts of a personal project.

First I pushed GPT-5 to clean up as much as it could of some horrific mess that had accumulated in my frontend test suite, then I threw some feature requests at it.

Anecdotally so far, GPT-5 seems quite good at adding features, following (and helping me revise) plans; fixing bugs. At least on par with sonnet 4, my go-to for the past few months; I think it's making fewer mistakes and getting stuck less often than anything else I've tried. It is rather chatty though.

This was not a clean experiment though, as the very first thing I asked GPT-5 to do was streamline my Cursor rules, which had gotten very bulky and verbose.)

What's the furor about Cursor?

For anyone not aware of the kerfuffle around Cursor's recent subscription plan changes, here's one of many articles.

TL;DR from my POV:

  • Cursor used to give you seemingly endless amounts of LLM usage for the $20/month "pro" flat rate , AND a discount for people who paid for a year up front;
  • Cursor recently made a series of confusing subscription plan changes that appear to walk that back;
  • Anecdotally, a lot of heavy "Pro" users suddenly started hitting their monthly quota within a few days of their monthly reset. And they were PISSED.

Here's a screenshot from the user forum digest in my inbox circa August 20th 2025:

screenshot of angry user comments

Anysphere, the company behind Cursor, was probably bleeding money at a prodigious rate prior to the announced changes. I haven't followed the details closely, but apparently they are the single largest user of Anthropic's models as of now (summer 2025), and I've read that Anthropic raised the prices they were charging Anysphere.

(Search for news with both Anthropic and Anysphere and you'll find there's other wild stuff going on. I guess that's inevitable when your biggest supplier becomes your biggest direct competitor overnight.)

I would guess that both companies probably saw little choice but to change something to staunch the bleeding. I've long heard commentary that there's an AI bubble built on unsustainable pricing, losing billions to gain market share, etc. It's pretty clear that's what is playing out here.

But regardless of necessity, it's clear that Anysphere fumbled the communication to their customers badly, and are now finding out the hard way that when you change what people think they have paid in advance for ... well, you can say whatever you like about how you never actually promised unlimited usage, but you will forever be seen as the jerk who took away the free doughnuts.

What Pro buys: Still clear as mud

I'm not much invested in how Anthropic vs Cursor plays out from a horse race perspective, but as a user of these tools, I don't have unlimited funds to throw at them, and I wanted a plan B in case it does turn out that Cursor gets prohibitively expensive for me.

This did prompt me to start exploring other ai coding tools (Aider, Crush, Claude Code...), which is probably a good thing for me regardless. I have drafts in progress about some of this and need to get more posts live!

So. Anecdotally my $20 monthly fee bought about $60 worth of usage before my last monthly reset, and I didn't hit any limits. Can I rely on that going forward? I have no idea, and I'm still using Cursor productively, but I'm keeping an eye on the exits.

screenshot of cursor usage dashboard showing $40 of usage and counting, 2 days before reset