Why are the models on using your own keys so much better?

General
1

Hi All.

I wonder if someone from cursor could shed some light on this.

Ive been using cursor paid for a while now. Ive noticed the models seem to have gotten increasingly more “stupid” with code questions especially when it comes to any larger chunks of code.

Using paid cursor I get these options

e9f570d8aa3177380268a8ab6a419151e06a8e7941f6195db4a699563766d288
e9f570d8aa3177380268a8ab6a419151e06a8e7941f6195db4a699563766d288454×576 17 KB

I always choose GPT-4 or cursor-fast which I am told is supposed to use the latest turbo models, is that correct?

But then if I swap to my own keys I can choose from

1a33bfc30d97c6b010e89257c04b2d745ebe6fa8bfce4467cba9ed63601550db
1a33bfc30d97c6b010e89257c04b2d745ebe6fa8bfce4467cba9ed63601550db500×516 20.9 KB

I’ve been running into an issue nthepast feqw days where the GPT4 and Cursor-Fast models were failing miserably. GPT just going in circles and failing to get to the answer.

I swap to GPT-4-1106-preview using my own keys and suddenly it nails the issue right away.

So my questions:

  1. Can someone shed some light on which exact models gpt-4 and cursor-fast under the cursor paid plans are using as this is not clear.

  2. Why is GPT-4-1106-preview so much better than gpt4 or cursor-fast? It seems counter productive to pay for cursor but then to have to keep swapping to my own keys (in turn paying Open AI) in order to get decent answers out.

9 Likes
2

I have a similar question:

For cursor chat, does

  1. cursor-fast use gpt-4-0613 or gpt-4-1106-preview or gpt-4-0125-preview
  2. GPT-4 slow requests uses 0613?
  3. What about inline edit (ctrl-k) which base model does it use?
3

Would be great to get some clarity on this! Its the source of much frustration daily for me and I still find consistently that cursor-fast or gpt4 using cursor is not as intelligent as the preview models using my own keys.

4

Do you have an example of where using your own key gives you better answers?

  1. cursor-fast: a 3.5-level model that follows our cmd-k format a little bit better
  2. gpt-4: since a few weeks ago, has been all gpt-4-0125-preview
1 Like
5

Could it be that when using your own API, your context length is 128k tokens but it’s only 8k when using cursor?

6

No, we’re using 10k tokens regardless of model.

7

Just to clarify, you use 10k tokens always? But GPT4 turbo supports 128,000 did I understand that correctly?

So do I understand that if I hit issues where cursor is struggling due to large chunks of code (where token size seems to be the cause) ill get better results going direct to https://platform.openai.com/playground?mode=chat&model=gpt-4-1106-preview and pasting in there?

Would be great to get clarification on this as I feel token issue seems to be a real issue with some of longer conversations with larger chunks of code. It feels “stupid” but actually just dont think it has the token reach to have all the context it needs.

1 Like
8

How can I use the 128k context in openai? never used my own keys, should try. I find the default model having a bit of hallucinations (suggesting non existent but ‘would make sense if it exists’ classes and functions as chatgpt does :sweat: ) I wonder if those other latest versions that work only through api keys are better, really want to try. should I configure it first on openai side for the limits of context window and then use my keys in cursor? (can I use my keys in Cursor free version as well? I’m still using the free)

1 Like
9

In the past week it has gotten so much worse! However, I’ve noticed ChatGPT is also suffering from serious comprehension issues.

cursor-fast was fairly decent, but now it’s completely unreliable, struggles with the most basic asks even with small context being fed into it.

gpt-3.5 the same issue, it’s having difficulties following directions

10

From what I understand, Cursor is currently using the GPT-4 0125 model, which, according to this article: URL, has become even worse. If you connect your own API key, you will have access to the older GPT-4 1106 model, which handles tasks better and hallucinates less. And yes, you can use your key without switching to a paid plan in Cursor.

2 Likes
11

So cursor is using only 10k token context window? Because I also have a ChatGPT enterprise account which seems to use128k token context window. So not sure if I should better use the enterprise acccount for some things?