Claude Sonnet 4.5 Seems Dumb this month?

Bug Reports
1

Where does the bug appear (feature/product)?

Cursor IDE

Describe the Bug

For some reason, sonnet 4.5 seems really dumb the last 2 weeks? Like it’s behaving like claude 3.7 at best???

Is this a bug or some shady things going on behind the scenes??

Steps to Reproduce

Claude 4.5 is making unnecessary changes elsewhere, like 3.7 used to do. Makes dumb changes. Forgets instructions…

Operating System

Linux

Current Cursor Version (Menu → About Cursor → Copy)

Version: 1.7.43 (system setup)
VSCode Version: 1.99.3
Commit: df279210b53cf4686036054b15400aa2fe06d6d0
Date: 2025-10-10T04:21:47.663Z
Electron: 34.5.8
Chromium: 132.0.6834.210
Node.js: 20.19.1
V8: 13.2.152.41-electron.0
OS: Windows_NT x64 10.0.26100

For AI issues: which model did you use?

Claude Sonnet 4.5

Does this stop you from using Cursor

Sometimes - I can sometimes use Cursor

1 Like
2

I think I also noticed this. Like when it was originally launched, it was great at doing big tasks in one shot. Now I feel like there’s no difference between the new haiku and that model, given the huge cost difference. I just use Sonnet 4.5 to tackle something complex which the auto mode or haiku fails to do. But overall, yes, I felt the same.

1 Like
3

I work with Sonnet 4.5 Thinking in planning mode and with regular 4.5 in Agent mode. I haven’t noticed any decrease in performance. But what is there is that it ignores instructions. I managed to get it to stop creating dozens of md files, but so far I haven’t been able to get it to stop printing a list of all the changes it made as part of the task.

1 Like
4

Yeah it was great when it launched, now it behaves more like Haiku 4.5.

6

Codex is the beast though this month.

7

How do you use it? It’s so slow that it’s beyond belief. I recently ran Codex and Sonnet for comparison. Codex ran for about 40 minutes with additional requests and failed to complete the task. Sonnet ran for 5 minutes and did everything right. So why wait 40 minutes when you can run Sonnet in 4 minutes and finish the task with prompts?

8

codex is the best token wasting model ever. GPT-5 high is better, Codex for some reason likes to stop all the time.
i did send it a task to do but it just listed all cursor tools for because i had a rule telling it that it has access tools.

sonnet 4.5 is lazy it will not do the tasks that you told it to at all it always do half implementation of it.