Sonnet 4.6 overthinks like crazy in cursor

Support Bug Reports
1

Where does the bug appear (feature/product)?

Cursor IDE

Describe the Bug

For tasks that I previously used Sonnet 4.5 and had no issue with, Sonnet 4.6 uses a ridiculous amount of thinking tokens and has a very strong tendency to overthink and use a large amount of reasoning tokens without actually producing a meaningful output. I did a direct comparison with Sonnet 4.5, which was able to smoothly reach the solution. Not sure how this was intended to be ingrated but I think the 4.6 series in general (both Sonnet and Opus) needs to be further investigated and refined because many users have already made complaints and observed how many thinking tokens these models use for certain tasks.

I understand that Anthropic now has a thinking effort parameter and is deprecating the budget tokens, so tuning is not completely in your control, but I think perhaps adjustments in the prompt or certain parameters could be fruitful.

Steps to Reproduce

Ask for it to make a simple scraper or something that collects data sources from different platforms and analyses using an LLM. This is just an example but it should be fairly straightforward to reproduce.

Expected Behavior

Produce at least some sort of output that adheres to my request and not have a large amount of reasoning tokens used for simple tasks

Operating System

MacOS

Version Information

Version: 2.4.27 (Universal)
VSCode Version: 1.105.1
Commit: 4f2b772756b8f609e1354b3063de282ccbe7a690
Date: 2026-01-31T21:24:58.143Z
Build Type: Stable
Release Track: Default
Electron: 39.2.7
Chromium: 142.0.7444.235
Node.js: 22.21.1
V8: 14.2.231.21-electron.0
OS: Darwin x64 22.6.0

For AI issues: which model did you use?

Sonnet 4.6

Does this stop you from using Cursor

Sometimes - I can sometimes use Cursor

3

Hey, thanks for the report.

To help us look into this, could you share a couple of things:

  1. The Request ID from a specific chat where Sonnet 4.6 “overthinks” (top-right of the chat > Copy Request ID)
  2. Were you using this in Agent mode or Ask mode?

Also, you’re on version 2.4.27, and we’ve shipped a few updates since then. It might be worth updating to the latest version to see if the behavior improves, although it’s probably more about a server-side setting.

In the meantime, if Sonnet 4.5 works well for your use case, it’s a solid fallback while we investigate the thinking-token behavior in 4.6.

4

Hey thanks so much for your quick reply!
So, i updated the version to latest and unfortunately this issue still exists, to be fair i blame this more on sonnet 4.6 extended thinking seems to be a bit wild. The request id is:

bc459e21-0cd6-442f-8ecb-15b0cee39cb6

Agreed, for now I’ll continue using sonnet 4.5 I do think 4.6 is actually better and more thorough but it’s eating up my usage. Please let me know if there is anything I can do from my side