I cannot pin the Grok 4.7 reasoning effort for child Cloud Agents in a Project.
I set the Cloud Agent model to Grok 4.7 Extra High (256k context, Fast off). Child agents still run as Grok 4.7 Low. The model slug that actually runs is grok-4.7, not grok-4.7-xhigh.
This happens in both of these cases:
The child is started with an explicit model slug of grok-4.7-xhigh.
The child is started with no model slug, so it should inherit the Project / Cloud Agent default.
The system prompt on those children only says “You are Grok 4.7”. It does not say Extra High. Cursor Usage bills them as Grok 4.7 Low.
The same Project can pin Grok 4.6 Extra High. A child started with cursor-grok-4.6-xhigh shows that slug and Usage matches. The bare slug grok-4.6-xhigh is rejected as invalid (Model 'grok-4.6-xhigh' is not available or invalid).
Changing the account-level Cloud Agent default, or the model picker on the already-open Project, does not rewrite the Project’s stored model. This Project’s originalModelName stays grok-4.7 after the picker is set to Grok 4.7 Extra High.
Steps to Reproduce
Open a Project Cloud Agent and set its model to Grok 4.7, Extra High, 256k, Fast off.
From that Project, start a child Cloud Agent with model slug grok-4.7-xhigh.
Start another child with no model slug.
Compare each child’s system-prompt identity and the Cursor Usage row.
Expected Behavior
Explicit grok-4.7-xhigh runs as Grok 4.7 Extra High, and Usage shows that effort.
Omitting the slug inherits the Project setting (Grok 4.7 Extra High), not a bare grok-4.7 Low default.
Hey @Hopeless, thanks for the detailed report and table.
You’re right about children started without an explicit model. In that path, the Project passes the Grok 4.7 model name but does not carry over the selected effort level. We’re tracking this issue.
For the two children we checked that were explicitly started with grok-4.7-xhigh, the requests were routed and billed as Extra High. As a workaround, have the Project explicitly pass grok-4.7-xhigh whenever it starts a child. You can reinforce this in the Project instructions with: “When starting agents, always set the model to grok-4.7-xhigh.”
Two details may explain the mismatch in the evidence:
You are Grok 4.7 identifies the model family, not the effort level, so it does not distinguish Low from Extra High.
When a child finishes, the Project can make a separate coordinator turn to read the result. That turn has its own Usage row, which can appear immediately after the child’s rows and show bare grok-4.7.
If you can reproduce an explicitly started grok-4.7-xhigh child appearing as Low, please share the child agent link. We’ll inspect that run directly.
Thanks for looking into this, and for separating the two cases.
The no-slug path matches what I saw: the child gets the Grok 4.7 name and not the selected effort. I also understand that “You are Grok 4.7” only names the model family, and that a later Project turn can be its own Usage row labeled bare grok-4.7. I am not using either of those as the child’s effort.
The case that still reproduces is an explicit grok-4.7-xhigh child whose own Usage row is Grok 4.7 Low. The original post omitted agent URLs, so the two explicit children you checked may not be the run that showed as Low on my side.
Please inspect this one. It is a short probe, started with the explicit slug grok-4.7-xhigh:
Started 2026-09-24 00:51 UTC
originalModelName is grok-4.7-xhigh
The task was English only: quote the model-identity lines and stop. It did not read the repo or run tests.
In Usage, this child’s row is still Grok 4.7 Low, not Extra High.
If that row is actually Extra High and I am reading a following coordinator row, it should be easy to point out on this short probe. On my Usage page, the probe itself is the Low row.
The issue where Project children started without a model lose the Project’s selected effort has been addressed on our side and will be rolled out soon if it hasn’t been already. Once it’s live, those children should run with the Project’s full Grok 4.7 selection (Extra High in your case) instead of the default.
Could you give it another try after that and let me know if either kind of child (no model, or explicit grok-4.7-xhigh) still shows up as Low?
Thanks for digging into this, and for the note on the no-model fix.
I re-ran the probe today (2026-10-09). My Cloud Agent default started as Grok 4.7 Extra High. I compared the model picker, run-info.originalModelName, and the Usage Model column.
Explicit grok-4.7-xhigh (bc-af7fcefe) is Extra High in the picker, grok-4.7-xhigh in run-info, and grok-4.7-xhigh in Usage. It does not show up as Low. Explicit low (bc-c3545893) and medium (bc-db13a5fd) match on all three sides.
Explicit high (bc-91e0edca) shows High in the picker and grok-4.7-high in run-info, but Usage records grok-4.7.
A child given the bare slug grok-4.7 (bc-c9306dad), and a child started with no model argument (bc-71da7290), both report originalModelName grok-4.7. While the Cloud Agent default was Extra High, both were grok-4.7-xhigh in Usage, and the bare-slug child’s picker stayed on Extra High.
I then changed the Cloud Agent default to Grok 4.7 High and started a new Cloud Agent on Auto. Usage recorded grok-4.7, not grok-4.7-high.
So an unspecified child is not stably inheriting the selected effort. The runtime slug stays bare grok-4.7. Usage matched Extra High only while that was the default, and did not match High after I switched the default to High.
Happy to retry the no-model case again once you confirm the fix is live.
Thanks for re-running the probe with that much detail.
Good news first: the fix for children started without a model is live. In your run, the child given bare grok-4.7 and the child started with no model both picked up the Project’s full selection (Extra High).
A couple of things explain the rest of what you saw:
Children inherit the Project’s own model setting, not the account-level Cloud Agent default. Changing the default doesn’t change an existing Project. To make unspecified children run at High, change the model in that Project’s picker (or pass grok-4.7-high explicitly).
When a child finishes, the Project wakes up to read its result. That Project turn currently shows up in Usage under the child’s row and without the effort level, which is where the plain “grok-4.7” rows under your explicit High child came from. On our side, the High child itself is recorded as grok-4.7-high, and the Project’s turns ran at Extra High. We’re tracking the Usage labeling and I’ve added your report to it.
The new agent you started on Auto ran on Auto, so Grok effort didn’t apply to it. The grok-4.7 row next to it was the Project’s turn reading its result.
If you see a child’s own rows (not the ones right after it finishes) labeled with a different effort than it was started with, send me the agent link and the time of the row and we’ll check that run.
I understand the no-model fix is live, and that those children picked up this Project’s Extra High selection. I’ll treat the Cloud Agent account default as separate from this Project’s picker.
The plain grok-4.7 rows under the explicit High child, and the one next to the Auto agent, were this Project waking up to read the result, not the child running without an effort level. The Auto agent itself ran on Auto.
If I see a child’s own Usage rows, not the ones right after it finishes, labeled with a different effort than it was started with, I’ll send you the agent link and the time of the row.