When I set my model to Claude Sonnet 3.5 (the previous version or the newest one from a couple days ago), it’s just using GPT 4o obviously you can ask it and it’ll say that but it’s not a hallucination because I can tell how the GPT responses are formatted differently.
I don’t know if it’s because the API errors from earlier today and it just hasn’t switched back or something, but I’m definitely getting charged to use a model that it’s not actually using. Can someone please help me?
Cursor has definitely cut the context on the new Sonnet 3.5; was having so much fun yesterday, now it’s a debugging day. 
@skyfriends As you mentioned, Anthropic had some issues around the release of the new Sonnet update, and there was a short period (around 5hrs) where some queries were getting redirected to 4o, in order to avoid hitting you with errors.
This has been resolved now that Anthropic is back to normal. Apologies for any issues you had with things not performing as you expected!
I would appreciate it , if there is more transparency and the user is given the option to manually switch instead of autonmtaically switching to an inferior model because that can lead to a huge decline in performance and user experience. I was also having alot of fun with the new model yesterday but i noticed today even as of 40 minutes ago the quality of output has dropped dramatically. Please i hope there are no cuts are done on the backend that would hamper quality. I dont mind paying more provided quality is not compromised.
Unfortunately, this was something outside of the Cursor team’s control, and the best we can do is reroute queries to other LLM’s in the meantime. We don’t host or run the LLMs ourselves, so it isn’t a case of cost or bugs in Cursor itself.
This kind of downtime with an LLM provider happens very very infrequently, and on the odd occasion it happens, it usually goes unnoticed by most of our users. However, for the few that are used to the performance of a certain model, it can unfortunately cause some disruption such as in this case.
@danperks I understand that our queries are being rerouted; but having an extra UX that tells that our prompt are not actually being used for ‘claude-3-5-sonnet’ or the intended model (and actually rerouted to ‘gpt-4o’, for example) would go a long way.
I believe this extra UX is actually in place, but is used for scheduled downtime or longer issues - this one was over quicker than the UI was enabled!