Cursor Multitask/Build in Parallel uses different model than chosen

Support Bug Reports
, , ,
1

Where does the bug appear (feature/product)?

Cursor IDE

Describe the Bug

This happens in cursor iOS as well.

Whenever I use a frontier model or any model for plan mode, and then try using mutlitask or click build in paralel with composer 2.5 fast selected, it keeps using the model of the original plan mode.

Steps to Reproduce

  1. Plan using different model, usually frontier models
  2. After plan is done, below the markdown.md, there is a model selection option, choose composer 2.5. THen click build in Parallel.
  3. Keep the chat model at the bottom as is, it would usually be the orginal frontier model uses for plan mode.
  4. See the models that are chosen using frontier models.

Expected Behavior

The models that subagents choose should be the one I chose below the plan markdown option instead of the original frontier model.

Operating System

MacOS

Version Information

Version: 3.9.16 (Universal)
VS Code Extension API: 1.105.1
Commit: 042b3c1a4c53f2c3808067f519fbfc67b72cad80
Date: 2026-06-27T06:41:01.941Z
Layout: editor
Build Type: Stable
Release Track: Default
Electron: 40.10.3
Chromium: 144.0.7559.236
Node.js: 24.15.0
V8: 14.4.258.32-electron.0
xterm.js: 6.1.0-beta.256
OS: Darwin arm64 25.2.0

For AI issues: which model did you use?

Frontier Model: GPT 5.5, Opus 4.8
Implementation Multitask model: Composer 2.5 fast

Additional Information

It’s annoying, since I’ve realized I’ve wasted tokens on implementing it

Does this stop you from using Cursor

No - Cursor works, but with this issue

6

Thanks for the report, @ZoellaB! This is a known issue that we’re currently tracking.

As a workaround: parallel/multitask children currently always inherit the model used in the chat, so you can set the main chat model to Composer 2.5 before clicking Build in Parallel, and the correct model should be used.

8

I think I remember posting about this for Cursor iOS, as well, and I tried to look through the posts to see similar bugs, but couldn’t.

I’ll keep that in mind

9

I experienced this yesterday. Using /multi-task, auto, I assigned the agent the task of acting as an orchestrator and parallel-developing 3 phases of my application. I noticed Sonnet 5 Medium was chosen for a number of the tasks. I was surprised to see a 10% use of API in my costs page given I ran it under auto mode.

10

Same here.

The main session was on grok-4.5 and for some reason a gpt-5.5-xhigh was selected for multitasking and this took a nice chunky 10% bite out of my API usage. Luckily i was on 20% only, but such behavior makes multitasking a risky endeavour.

RequestId: 8464632b-4d54-491f-9027-30aa24d13e56