Grok-3-mini not usable

Bug Reports
1

Describe the Bug

Hi,
grok-3-mini quite often shows this behaviour:

  • messes up tool calls
  • runs circles updating todos instead of doing anything

I would be so nice to be able to use other models than anthropic but whatever you currently try, you’ll hit a wall if it’s not anthropic.

Like, seriously guys, Opus 4.1. barely made it through the door and you integrated it perfectly while all potential alternatives are not working and takes weeks for fixing.
How? Why?

Steps to Reproduce

use grok-3-mini

Expected Behavior

get tools calls right, do not run endless circles

Operating System

MacOS

Current Cursor Version (Menu → About Cursor → Copy)

Version: 1.4.2 (Universal)
VSCode Version: 1.99.3
Commit: 07aa3b4519da4feab4761c58da3eeedd253a1670
Date: 2025-08-06T19:23:39.081Z
Electron: 34.5.1
Chromium: 132.0.6834.210
Node.js: 20.19.0
V8: 13.2.152.41-electron.0
OS: Darwin arm64 24.6.0

2

Hi @proteus-dev some models are performing better while others may not have the same capabilities.

Mini models are usually less performant and may have limitations with Agent usage compared to regular sized model due to lower intelligence.

That explains also well why Opus 4.1 one of the best models, until today when GPT-5 was released, performs so well.

While we do testing and improve from our side the prompts, behaviors and tool calls there are still limits as we can not re-train the models.

3

Hey condor,

I have some doubts about this being related to intelligence.

This is about
a) getting tool calls right. An issue well-known, even with higher performing models like gemini 2.5 pro.
b) proper instructions for endless cycles detection. An issue well-known, even with higher performing models, like sonnet 4 or opus 4.

Moreover, this is about forum mods telling users to use lower cost models, which I would like to, but if simple things like a tool call (even GPT 3.5 was able to do a tool call) or avoiding endless cycles (this can happen with ANY model but needs to be detected and intervened) do not work, proposals regarding using lower cost models are not helping, simply because there are no alternatives.

Thanks for advertising GPT-5.