The model settings in chat (e.g. whether to use thinking and effort level) are getting reset to thinking high after every restart. Caused me to loose a lot of tokens every time. This has been mentioned before in another context, but now it’s just annoying that this happens after every restart, with every new window I open, etc.
It’s quite annoying how many times Cursor has completely disregarded user settings, this is one of them.
Steps to Reproduce
Just restart or update Cursor, open chat and be surprised how many tokens it uses because it’s been reset to thinking high once again…
Expected Behavior
I expect the cursor team to not throw out any good UI patterns, disregard user settings and reset settings over and over again (with this bug and other bugs this has been surely the 10th time I have to adjust my settings on every restart/update).
This is a known class of bugs where model settings like thinking/effort persist across restarts. Some fixes are already in 3.1.17, but the newest one from April 20th, about the case where a second open Cursor window brings back settings you changed in the first window, hasn’t made it into your build yet. It should land in the next update.
A couple quick questions to narrow this down:
Do you usually keep multiple Cursor windows open at the same time? The latest fix is specifically for that case.
Which model is selected (for example, Opus 4.7), and what do you see after restart? Is it the same model but forced into thinking, or does the model itself also change?
Yeah I almost always work with multiple windows. I mostly use Opus 4.6 and Sonnet 4.6 without thinking and in medium effort. After a restart/opening new windows it adds the thinking and high effort. It only happens sometimes, but this has been going on as you say for weeks if not months with a back and forth on how we choose models and where we do the settings for thinking/non-thinking and effort.
The bigger issue here is not really the bug, which is most annoying in itself, but the absolute inconsistency, constant changes to the UI that are absolutely unwarrented, outright confusing and as we can see in this case harmful to users wallets if they don’t pay close attention to an 8-pixel icon that keeps changing to the most expensive mode.
Yeah, multi-window is exactly the case that the April fix addresses. The second window was bringing back settings you changed in the first one. It’s not in 3.1.17 yet, it should land in the next update. If it still resets to thinking high on Opus or Sonnet 4.6 after updating, reply here and we’ll dig in.
On the UI side, I agree. The model picker has been changed too often over the last few months, and switching models changing the request cost without a clear signal is a real issue. I’ll pass that feedback to the team separately from the bug.