Does gpt-4 model in AI chat use gpt-4-turbo?

I have read in some posts that it does but I couldn’t find any official post announcing this. Can someone confirm, please?


1 Like

This used to be true but I believe the creators of Cursor intentionally rolled back to GPT-4 regular as they had issues with hallucinations and laziness in GPT-4 turbo.


Chat uses gpt-4-turbo right now.

cmd-k uses gpt-4-turbo, but we may be rolling back due to the laziness problem. Chat will stay unaltered though

If you do roll it back, can you please ensure there is a clear way to select turbo when needed (the token size is a huge difference).

It would be great to have one place to check to see what setting in cursor is currently using which model.

Love cursor but this is an issue i think frustrates many it feels a bit of a “black box”. More transparency here I think would be appreciated by everyone.


@jasondainter I think CMD-K has been rolled back now. However, in the Cursor settings popup under “OpenAI API”, you can add a model called gpt-4-0125-preview and use that model in your CMD-K queries. This should force the new Turbo model. And that will work even if you don’t use your own OpenAI API Key.


Oh ok cool thanks, I didnt know those settings worked without an API key enabled. Are you sure that is the case?

@jasondainter Only the gpt-4-0125-preview model in there works without an API key. But it will count as one use of your Cursor’s subscription GPT-4 usage. And I think you need to have a Pro subscription. The GPT-4 uses from the free trial don’t work with it.

Ho God, this explain why suggestions is so lazy and unusable this days.

what does “laziness” means in this context, slow response from the model?
btw I want to ask if it’s possible to specify a fine-tuned gpt3.5 model from openai in case I make a fine tuned one for the DSL I’m working with (which is not known to gpt4 since it is a completely new thing and not in any data sets) ? Is it simply using the correct name of the fine-tuned model from my account and my key, would that work? (or only public model names can be used, but not customized ones?)