API/Model selection is completly brken

Bug Reports
1

Model selection (via API) stop working reliably. I have found a couple bugs:

  1. When overriding OpenAI base URL (e.g. to Deepseek or Openrouter) and then trying switching back to the default one via the reset button doesn’t work, it keeps populating the field with whatever you put there previously. One has to not only press the reset button but also the save button several times for it to stop doing this.

  2. Google API doesn’t work anymore. Cursor always tries to reach OpenAI and fails with reason: incorrect API key provided (probably because it tries to use Google’s key). This doesn’t happen with Anthropic’s.

  3. When selecting OpenAI API and o1 (or o1-2024-12-17) as model, key verification fails with reason: unsupported parameter max_tokens… one has to first select gtp-4 in order to first have a successful key verification and then being able to select and use o1.

The whole “Models” section really need an urgent revamp! Especially disentangling OpenAI stuff from everything else

I am using version 0.44.11 on windows

2

If you override your OpenAI URL, we don’t know which models you want redirecting through there, so Cursor will attempt to use all models over your custom URL.

We are aware of the o1 issue with verification too.

We are working on some improvements around this internally, and hope to ship these soon to help improve the clarity and stability of custom API usage.

3

Yes, but even with the default OpenAI URL, accessing Google models via the Google API stopped working. Before I was able to reliably use OpenAI, Google and Anthropic models. Why is it trying to use reach OpenAI when using Google models?. This happened after I modified the OpenAI base URL (and then reverted to the default one)

Edit:

Forgot to add that using the “verify” button to verify Google’s API key works without a problem. Also, these bugs are both present in chat and with crtl + k mode

4

If you remove all your API keys, and overrides, do the Gemini models work in Cursor with just your Pro plan?

5

Yes it does, but at the same time, if I enable only the API for Gemini I get the “unable to reach OpenAI” error. The same doesn’t happen with Anthropic. I am now on 0.45.4. I depend on keys because I go through my fast requests quite quickly, so not being able to use my own keys is quite a downer

6

Hey, I believe one of the Gemini models, gemini-exp-1206 is no longer functional, so I’d recommend disabling that model to ensure that isn’t being used for verification or prompts.

If that doesn’t help, let me know and we can investigate further!

Just to let you know, we have improved our ‘usage pricing’ feature, so that instead of needing API keys, you can enable that and be billing $0.04 per fast request, with no minimum or maximum usage. You can set a spend limit in the dashboard, so you don’t spend more than you want, but this should be easier and simpler than managing API keys, without having to commit to batches of 500 like we originally offered.