Hey @Scott_Heffield, thanks for the detailed and calm message, it really helps.
Here’s the status with no sugarcoating. The main agent is still running on the Grok model you selected, but some helper and subagent tasks (computer-use, browser, video review, exploration) still pick a model on their own and don’t inherit the parent model. That’s why you’re seeing usage from the Other Models pool. It’s not something in your setup.
Right now there’s no reliable way to pin all subagents to your selected model. Settings → Models only limits the picker in the IDE and doesn’t affect cloud helpers, and the instruction “don’t use subagents” works inconsistently. I don’t want to suggest a workaround that doesn’t actually work.
We’re tracking this issue and this thread is linked to it. I’ll pass your point along that this hits power users the hardest, both to Product and Support. I can’t share an ETA yet, but as soon as there’s real progress, I’ll update you right here.
About usage charges for model usage you didn’t choose. It’s best to review this on your specific account. Email [email protected] with details (usage for the period you care about, screenshots) and the team can take a close look at your case.
@deanrie Just today, I’ve been building with cloud agents, and they are explicitly told not to use third party models. I can’t find any third party usage in my usage list for today. Yet my third party usage graph has gone from about 70% to 98% in an hour of building. This means that if I keep building this month I will pay overage costs.
@Scott_Heffield Cloud agents burn Other Models, I believe. Separately from whatever underlying model they’re using (even when using a Grok model). Staff please correct me if I’m wrong, I’d also like a concrete answer; but that is my findings.
You should setup self-hosted cloud agents. I posted a summary of my findings here:
If you want a detailed setup guide and the scripts you need to run self-hosted cloud agents, let me know, because I was debating publishing mine.
You could get by with Cursor’s docs, but mine should be easier to follow. You’ll need the ability to create a virtual machine, if you want to use my guide & scripts… but it can be done without one.
Self-hosted cloud agents means that you can run an unlimited number of parallel cloud agents, for an unlimited amount of time; without burning any Other Models usage whatsoever (you’d then only be burning your Cursor Models pool - for the actual AI usage, not for the cloud agents usage).
Hey @DannyB, thanks for trying to break this down for everyone.
Let me clarify the mechanics so we don’t lock in the wrong understanding. It’s not that cloud agents always spend Other Models on their own, regardless of the model you picked. The main agent runs on the Grok model you selected. But some helper and subagent work (computer-use, browser, video review, exploration) currently chooses its own model and doesn’t inherit the parent model. That’s what can cause usage from the Other Models pool. We’re tracking this, and this thread is tied to that issue.
I won’t give a hard rule for which activity maps to which pool. The mapping can change, and the current breakdown for your specific account is always visible in Dashboard > Usage.
On self-hosted: it can be a good option if you want tool execution to run on your machine, but be careful with the conclusion that it fully removes Other Models usage. With self-hosted, tool execution moves to your machine (file edits, terminal, computer-use, MCP), but the agent loop and inference still run on Cursor’s side, so model usage is still counted. I can’t confirm that this means Other Models won’t be used. If you want to test it, compare Dashboard > Usage before and after, that’s the source of truth.
For charges tied to models you didn’t pick, it’s best to have support look at your specific account via [email protected] (sounds like you already emailed, that’s the right path).
No ETA on the main issue yet, but I’ll reply here as soon as I have an update.
I’ve read this thread, and just want to add that this is a very frustrating problem. It seems to me that all Computer Use subagents are pinned to Sonnet 4.5, which is a very old model at this point. So it’s not very good at it’s job… It’s very confusing why the main cloud agent is prohibited from driving the computer use function, or why we can’t have computer use subagents that use anything other than Sonnet 4.5.
I would definitely appreciate this issue getting prioritized! It’s enough to make me want to never use cloud agents, or at least never use them to click through UIs that they built. Which is super unfortunate because it’s a very powerful pattern.
Workaround that works for me (slower, saves tokens)
Same issue: Cloud Agent on Grok, then computer/browser helpers still burn Sonnet even with Claude disabled in Settings → Models.
What I do:
Cloud Agents only write code and open/update the PR. Prompt them hard: no browser, no computer use, no UI screenshots via computer subagents.
Force the launch model (Grok) with Fast off.
UI / click-through QA runs in Grok Bot on its own computer after the PR lands — that uses Grok Bot tokens, not Cursor Cloud Sonnet.
Order: diff review first; Grok Bot browser QA only after there are no major change requests left.
Slower than Cloud Agent click-around mid-run, but it stops Sonnet/Fast eating quota on trivial verify. Settings → Models still doesn’t block those helpers, so this is the control I have until inherit is real.
Same issue here, and honestly it’s really frustrating. My agents are set to Grok, but the computer-use helper keeps running on Claude Sonnet 4.5 anyway. This cycle it’s used almost 100 million Sonnet 4.5 tokens, about 35% of my extra usage, just on clicking around to check UI changes. Turning Claude off in Settings > Models does nothing.
I don’t buy that Sonnet 4.5 is the best model for computer use anymore, either. It’s old, it’s expensive, and newer models do this as well or better. I’ve had to stop my cloud agents from using the browser at all just to stop the bleeding. This thread has been open a long time, and it’s hard to believe it still isn’t fixed.
Hi @deanrie I just wanted to check on updates here. With the new offer of combined billing for Cursor + Grok with the X subscription plans, which removes allowance for third party models (all billed at API rate), this becomes even more critical to anyone wanting to use Cloud Agents. I’m holding off on converting billing, just for that reason. So any bone you can throw on prioritization for the issue, if it’s being actively worked, etc. would be appreciated. Thanks in advance!
Adding my voice again, almost two months in. This thread now has dozens of paying users reporting the same thing: we pick Grok, disable Claude, and helper or subagent tasks still drain our Other Models quota on Sonnet 4.5. Several people here are cancelling or holding off on plan changes because of it. We’ve had repeated explanations of the cause, but no fix, no ETA, and no way to opt out. Please give this real priority and share a timeline. At minimum, let us turn third-party helper models off, or refund the usage they consumed without our consent.
I encountered the same issue as well and hope it can be fixed quickly.