The reason why most people have negative experience is, I believe, because gpt-5 appears to interpret prompts completely different than previous models did.
In this particular case you sent, the user asked the agent to “please suggest changes”
And it implemented it. I have successfuly and consistantly made it to only suggest changes by explicitly prompting it “Do not edit anything yet, only give me a plan of how you will implement this”
gpt-5 is not inherently bad, it’s behaviour is just completely different, and requires more precise and literal prompting - which will give you more precise and higher quality results.
My model list still hasn’t updated. I manually tried configuring an OpenAI key, but the main issue is I don’t know where to get this key. If I were to use my own OpenAI key, I assume it would incur charges, right? Currently, if GPT-5 appears automatically in the model list, it should still be free to use.
This is a great assessment, GPT5 feels like the reliable and by-the-book whilst sonnet-4-thinking which is my go to feels more like the daring rockstar. I’m going to be trialling using both in tandem to match the problem at hand.
You can get an openai key from https://platform.openai.com/api-keys but I don’t think that will mean that it’s free to use if you’re bringing your own key.
Have you updated to the latest version of cursor? The models should appear there. You may need to go into the settings → models and then enable them in the list of cursor models
Im using both right now, gpt-5 for the grunt work, opus to check its work . Opus is more expensive, they need to drop it down, but it hands down produces better code in my (few days) experience. Talking to GPT-5 is also like talking to a rock sometimes. You go back and forth like 5 times and it still doesnt get it, then I switch to claude and go “Hey man can you sort this out” and it produces a perfect summary of what is wanted and the right implementation :D. Definitely think Claude is ahead of GPT-5 in many many ways. If they can do another iteration that keeps it from always just randomly writing code when asking questions and be more effective on cost, I’d not be using 5 often. (5 does seem to do a good job of digging deep and sussing hard problems out, I just dont trust any of its solutions).
As an ardent Cursor supporter, I do wish we had more clarity on communication when it comes to a simple usage / tracking system for all of this.
The GPT5 models are still marked free in the app, the usage cost has been tracking the entire time (the GPT5 trial time) in accounts–meaning very shortly we will not know what was included in the “free” portion, and what we’ll be charged for.
I understand, as did many people, the need to change the pricing, but adding ambiguity to the system has introduced a pain point that can be simply avoided by providing an abundance of clarity.
Hey @Deebee50 GPT5 models should no longer be marked as free in the app (they’re currently being charged per this post). Usage during the promotional period was all free, however!
Ya i mean it’s a fair question - at least we got some official clarification though. Hopefully that checks out. It is hard to tell if the delta above “at least $60 worth of api tokens” would be higher or lower without the GPT5 spend, but that’s the nature of it.
I think “included in pro” was catching you off guard but probably they just didn’t have another way to classify it in the raw events list (more trouble than it’s worth to change just for this week). Might be better if it said “Free bonus” instead?
For ultra plan members, will the token usage from the GPT-5 models during the free promotional period be shown? There is currently no indication on the dashboard that any portion of the usage is free.
Correct, there is nothing in the dashboard indicating it was free. However, in the application if you have the usage visible there (enabled under chat settings), then you will see the true usage, and it will most likely show that the gpt-5 usage during the promotion is not counting towards your limits. Why they can’t make the usage page on the web dashboard more clear is ridiculous, and I think it can only be to create intention obfuscation for when they make the next price change.
. Opus is more expensive, they need to drop it down, but it hands down produces better code in my (few days) experience. Talking to GPT-5 is also like talking to a rock sometimes. You go back and forth like 5 times and it still doesnt get it, then I switch to claude and go “Hey man can you sort this out” and it produces a perfect summary of what is wanted and the right implementation :D. Definitely think Claude is ahead of GPT-5 in many many ways. If they can do another iteration that keeps it from always just randomly writing code when asking questions and be more effective on cost, I’d not be using 5 often. (5 does seem to do a good job of digging deep and sussing hard problems out, I just dont trust any of its solutions).