Cursor Projects agent used an unauthorized model

Where does the bug appear (feature/product)?

Cursor Agents Window

Describe the Bug

I have disabled all models except Composer 2.5 and Grok 4.6. My cloud agents’ default model is Grok 4.6 on Medium.

I asked my Cursor Projects orchestrator agent to do some work. It spawned an agent which ended up using Claude Sonnet 5 on High thinking. This ended up consuming my “Other Models” quota. Note that Claude Sonnet 5 was disabled on my end.

I love Cursor. But Cursor has a major issue that’s been going on for a long time now where it switches models/modes unexpectedly. Sometimes, it’s using Fast mode even when I have explicitly disabled it, defaulting to Grok over Composer, not using the defaults, not respecting inherit mode for subagents, and now using an unauthorized model that was explicitly disabled.

Some of these may have been fixed, but there’s a common pattern around incorrect model use by Cursor. Please fix this. The products and models are great otherwise.

I asked the orchestrator agent why it used Claude Sonnet 5 High, and here’s what it said. It doesn’t have access to the list of authorized models.

Steps to Reproduce

You can debug my chat sessions from Sep 11, 07:09 PM UTC to Sep 11, 07:26 PM UTC. Most of them used Grok 4.6 Medium (orchestrator model) but the worker agent used Claude Sonnet 5 High which consumed 6.2M tokens.

Expected Behavior

Only authorized models are used.

Separately from this ticket but related issue, Fast mode isn’t used ever if not enabled. Maybe give us a way to permanently opt out of Fast mode.

Operating System

MacOS

Version Information

Version: 3.20.10
VS Code Extension API: 1.128.0
Commit: d6f462cdd0a6a6d1cff570daf980e671d0a63de0
Date: 2026-09-10T21:04:54.045Z
Layout: Agent Window
Build Type: Stable
Release Track: Default
Electron: 42.10.0
Chromium: 148.0.7778.280
Node.js: 24.18.1
V8: 14.8.178.38-electron.0
xterm.js: 6.1.0-beta.291
OS: Darwin arm64 25.5.0

Does this stop you from using Cursor

No - Cursor works, but with this issue

Hey Archit, thanks for the detailed report and the screenshot. I can see where the orchestrator says it doesn’t have a list of authorized models.

I checked what happened. Your orchestrator ran on Grok 4.6 (Medium), like you selected. But the worker agent it created started on Claude Sonnet 5 because the orchestrator set that model for the worker, even though you didn’t request it. That’s what used your included usage.

A couple notes on how this works right now:

  • The list in Settings → Models only filters the model picker in the app. It doesn’t yet limit which models cloud workers and helper agents can run on.
  • By default, the worker should inherit the orchestrator’s model, so Grok 4.6 should have been used for the whole chain.

This is a known issue we’re tracking, same as here: Cloud Agent logs claude-4.5-sonnet when picker is Grok 4.6 and Claude is not enabled. It’s definitely not anything in your setup. I can’t share a timeline yet, but if there’s an update I’ll reply in the thread.

A workaround that helps a lot is adding an instruction in your prompt or project rules like: “When creating worker agents, never specify a model. Always let them inherit my model.” It won’t be a hard guarantee, but it helps consistently.

If you want us to look into the specific usage that went to that worker, email [email protected] and they can check your account.

About Fast mode, that’s a separate issue, I’ll pass it to the team as feedback.

Thanks @deanrie