It’s switching to inferior models - how do I stop this? When I pick a model, I expect it to use the model I picked. Otherwise, whats the point in picking?
Hey. From the screenshot, it looks like it’s not the model itself changing. Cursor is delegating part of the work (that “Explore identity invite flow” step) to a built-in subagent, which uses Composer 2 by default for speed and cost savings. The GPT-5.5 you selected stays as the main model in the chat, and the subagent just returns its result to it.
This is expected behavior, but I get how it can look like a swap. Two ways to avoid it:
- Turn on Max Mode for your selected model. Then subagents inherit the parent model instead of using Composer 2.
- Create a custom subagent in
.cursor/agents/<name>.mdwith this frontmatter:
Then exploration or research tasks will run through the model you set.--- model: gpt-5.5 ---
There’s a lot of feedback on this forum thread with discussion: Subtasks use composer 1.5. The team knows the current default is confusing. Changes to the default behavior are on the radar, but there’s no ETA yet.
I’m fully aware this is a subagent, and the core issue still stands. There’s a clear expectation that the model I selected is the one being used. Instead, parts of the workflow are silently delegated to a weaker model, which leads to lower-quality context gathering and degrades overall agent performance. This directly impacts output quality.
This behavior downgrades the agent across the board. Cost tradeoffs should be decided by the user, not imposed implicitly by Cursor. I should not need to enable Max Mode just to ensure the system uses the model I explicitly chose. Creating custom subagents is also not a real solution. It adds unnecessary friction for something that should be the default behavior.
The default does not respect the user’s selection. It simply does not use the model that was chosen.
It is also reasonable to question whether this default is intentionally designed to reduce operational costs for Cursor as a company at the expense of user agent performance and output quality. How else would you explain this insane default?
Thanks, that feedback is fair and heard. Subagent delegation to Composer 2 by default, while a different model is selected in the main chat, is a real source of friction, and there are plenty of forum complaints about it. We’re discussing this internally, and the current default isn’t considered the final decision.
A couple points I want to clarify, without trying to justify the design:
- Max Mode and custom subagents aren’t the real solution, they’re a workaround, and I’m not trying to present them otherwise. Right now, they’re the only ways to lock the selected model at every level. I get that this shouldn’t be on the user.
- The main chat model isn’t being swapped. What’s being swapped is the helper task model like exploration and search. Not an excuse, just to make it clear where the issue actually is and where it’s being routed.
On changing the default so it inherits the selected model out of the box, that’s logged as an open item in our internal tracker. I can’t give a timeline or guarantees, and I don’t want to claim we’ll fix it soon if I can’t back that up. But reports with specifics about context gathering quality getting worse do help with prioritization. That kind of evidence is stronger than a general bring back model choice.
If you notice specific cases where Composer 2 in a subagent clearly breaks the result, please share them here with the Request ID. Those examples are genuinely useful when we discuss changing the default.