Describe the Bug
Frequently when using Claude Sonnet 4 thinking, the model will not used extended thinking at all, and will instead just answer immediately. It performs very poorly in these instances, and seems to generally lack chat context. For instance, in a recent query, the non-thinking Claude Sonnet 4 fixated on an unrelated issue (that I believe was mentioned in a recent chat summary in the attached context) and rapidly made several edits trying to fix a non-existent issue that was not mentioned in my query.
Usually, restarting cursor fixes the issue temporarily (for 1-2 subsequent queries). The issue is very frequent, and once it switches to non-thinking it never switches back until I restart the client. Possibly an issue with too much context?
Steps to Reproduce
Use Claude Sonnet 4. Query repeatedly. The issue seems to be more likely to occur if a single chat goes on for a long time. Eventually, the model stops using the thinking tool, which is apparent because it answers immediately and there is no “thinking” output at any point in the response.
Expected Behavior
The thinking version of the model should always use the extended thinking tool as its first action in order to consider the query and related context.
Screenshots / Screen Recordings
Operating System
Windows 10/11
Current Cursor Version (Menu → About Cursor → Copy)
Version: 1.1.3 (user setup)
VSCode Version: 1.96.2
Commit: 979ba33804ac150108481c14e0b5cb970bda3260
Date: 2025-06-15T06:55:04.603Z
Electron: 34.5.1
Chromium: 132.0.6834.210
Node.js: 20.19.0
V8: 13.2.152.41-electron.0
OS: Windows_NT x64 10.0.19045
Additional Information
It’s a fairly recent issue. Earliest occurrence I found was 4 days ago (6/12/25), but it has been happening constantly since then.
Does this stop you from using Cursor
No - Cursor works, but with this issue