Limited number of models for cloud subagents

Why there is limited number of models available to be spawn as cloud subagents? Here is the list that is available for me

Available subagent models

  • inherit
  • claude-opus-5-thinking-high
  • claude-opus-5-thinking-high-fast
  • claude-sonnet-5-thinking-high
  • claude-sonnet-5-thinking-xhigh
  • composer-2.5
  • composer-2.5-fast
  • cursor-grok-4.5-high
  • cursor-grok-4.5-high-fast
  • cursor-grok-4.6-high-fast
  • gemini-3.7-flash-high
  • gpt-5.6-luna-high
  • gpt-5.6-sol-high
  • gpt-5.6-sol-high-fast
  • gpt-5.6-sol-xhigh
  • gpt-5.6-sol-xhigh-fast

I don’t understand why you offer cursor-grok 4.6 high fast, but not regular high or even xhigh… The same comes for GPT5.6, there is no terra or even sol medium

Hey @Marcin_Miler!

We’re currently working on making the full model catalog available for use as subagents in the cloud, rather than relying on a hardcoded list. We don’t have an ETA to share just yet but I’ll keep you posted!

Thanks for the info! It’s very important for me to keep using cloud agents (I love cursor cloud!), as for now due to this limitation I need to switch to local agents sadly.

Hey @Marcin_Miler! Quick update on this, we’ve already expanded the set of models available to cloud subagents since you posted (GPT-5.6 Luna, GPT-5.5, Kimi K3, and several Claude additions are now in the list). Exposing the full model catalog still a WIP.

It’s still the same list as before

Slug Notes
inherit Default — uses the parent agent’s model
claude-opus-5-thinking-high Claude Opus 5, thinking mode
claude-opus-5-thinking-high-fast Claude Opus 5, thinking, faster variant
claude-sonnet-5-thinking-high Claude Sonnet 5, thinking mode
claude-sonnet-5-thinking-xhigh Claude Sonnet 5, extended thinking
composer-2.5 Composer 2.5
composer-2.5-fast Composer 2.5, faster variant
cursor-grok-4.5-high Grok 4.5
cursor-grok-4.5-high-fast Grok 4.5, faster variant
cursor-grok-4.6-high-fast Grok 4.6, faster variant
gemini-3.7-flash-high Gemini 3.7 Flash
gpt-5.6-luna-high GPT 5.6 Luna
gpt-5.6-sol-high GPT 5.6 Sol
gpt-5.6-sol-high-fast GPT 5.6 Sol, faster variant
gpt-5.6-sol-xhigh GPT 5.6 Sol, extended
gpt-5.6-sol-xhigh-fast GPT 5.6 Sol, extended + faster

Hey @Marcin_Miler, you’re right, I double-checked and your list is indeed unchanged. The additions I mentioned are rolling out gradually and haven’t reached your account yet; apologies for the confusion there.

Where does the bug appear (feature/product)?

Cloud Agent (GitHub, Slack, Web, Linear)

Describe the Bug

TL;DR
I write plans that assign each task a subagent and an exact model slug from the models I have enabled for chat. When I ask the parent agent to run that plan, Task often cannot take the slug the plan named, because Cursor bakes a smaller, frozen model enum into the tool schema. Chat-allowed should mean Task-allowed. If the plan says Grok (or any other enabled model), the subagent must run that model — not inherit, not substitute, not stop because the enum omitted it.

This is all in a remote-control cloud-agent connected to a linux machine; the IDE is in Windows 11

Detail
I write implementation plans that split work into a task DAG. Each task names a subagent and an exact model slug. Those slugs are models I have enabled and that appear on the chat allowed list. When I later ask the parent agent to execute that plan with subagents, the Task tool often cannot take the model the plan recorded, because Cursor injects a smaller, frozen model enum into the agent’s tool schema.

That breaks the contract I am trying to enforce: plan says model X → Task must run model X, as long as X is on my enabled/allowed chat models. I am not asking for hidden models, unpaid models, or admin-blocked models. I am asking to use, on a subagent, the same models I am already allowed to select for chat.

What I am doing

Enable models in Settings → Models (they show in the chat picker).
In a plan, each task has a field like Model: cursor-grok-4.5-high (or another enabled slug). The plan also says: never substitute a different model for a task.
I ask the parent agent to implement the plan using subagents according to the DAG, passing that slug into Task.
The parent’s Task schema was baked at session start and lists only a subset, for example:
inherit
claude-fable-5-thinking-high
composer-2.5
gpt-5.6-luna-high
gpt-5.6-sol-high
gpt-5.6-sol-xhigh
Grok 4.5 is enabled for chat in this account, but not in that enum. The parent is instructed not to pass a slug that is not listed. So it cannot execute the plan as written: it must omit model, substitute a listed slug, or stop.
Why this is wrong

The plan is the source of truth for which model runs which task. Chat availability is the source of truth for whether I am allowed to use that model. Task’s enum is a third, narrower list that neither the plan nor Settings → Models describes. That third list is what actually runs.

Consequences:

I cannot pin Grok (or any other chat-enabled model missing from the enum) on a DAG task and have a subagent honor it.
Changing Settings mid-session does not refresh the schema; even a new session may still expose only one variant per family.
There is no “subagent allowed models” setting. Workarounds (inherit, custom agent frontmatter, disabling other models) are not “run the model the plan named.”
Cursor staff have already described this behavior: Task slugs “mirror the model picker with a single variant per family”; exposing all variants on Task is a “known limitation”; a broader subagent allowlist “doesn’t exist yet.”

Requested fix

If a model is on my chat allowed/enabled list, it must be a valid Task model slug, including the effort/fast variant I named in the plan.

Parity: chat-enabled model ⇒ Task-accepted slug. No silent subset.
Plan execution: the parent must be able to pass the exact slug from the plan. If the slug is invalid, error; do not omit it from the schema so the agent cannot even attempt it.
Snapshot: either refresh the Task enum when Settings → Models changes, or show in the UI that subagent models are a session snapshot and how to rebuild it.
Control: if you keep a filter, make it an explicit setting defaulting to all chat-enabled models.
I am not asking you to pick the model for me. I already did that in the plan, from the allowed list. Task has to accept that choice.

Related:

Subagent Allowed Models
Task accepts Grok 4.5 Low Fast but executes Composer 2.5

Steps to Reproduce

In Cursor Settings → Models, enable Grok 4.5 (or any model you intend to pin). Confirm it appears in the chat model picker and that you can run the parent agent on it.
Leave other models enabled as well (Composer, Claude, GPT, etc.). Do not hide everything except Grok — the point is mixed chat availability.
Start a new Agent chat with a parent model that is not Grok (for example Composer 2.5 or Grok 4.6).
In that chat, ask the agent to inspect the Task/subagent tool schema and list every value accepted by the model parameter.
Observe that the enum is a subset of chat-enabled models. In this session it was:
inherit
claude-fable-5-thinking-high
composer-2.5
gpt-5.6-luna-high
gpt-5.6-sol-high
gpt-5.6-sol-xhigh
Grok 4.5 was enabled for chat and absent from this list.
Write a plan (for example .cursor/plans/example.plan.md) whose DAG has a task like:
T1
Model: cursor-grok-4.5-high (or the exact Grok slug shown in your picker)
plus: never substitute a different model for a task.
In the same Agent chat, ask: implement this plan using subagents according to the DAG, and pass each task’s recorded model slug into Task.
Expected: Task launches T1 on cursor-grok-4.5-high.
Actual: the parent cannot pass that slug because it is not in the baked enum. It must omit model, substitute a listed slug, or stop. Changing Settings → Models in this chat does not add Grok to the enum.
Optional checks:

Repeat step 3–5 in a new session after setting the picker to Grok 4.5 High. The Task enum may still omit Grok or only include one variant per family.
Confirm there is no Settings control for “models allowed on subagents.”

Expected Behavior

Should run the subtasks with the selected models from the allowed list of models (e.g. grok-4.6)

Operating System

Linux

Version Information

Version: 3.17.19 (system setup)
VS Code Extension API: 1.128.0
Commit: ae3a2b7231dd56194447fe4570dfdc61640b1e90
Date: 2026-08-24T06:42:14.583Z
Layout: IDE
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.291
OS: Windows_NT x64 10.0.26200

Does this stop you from using Cursor

No - Cursor works, but with this issue

Thanks for moving my report here.

My list of models available for subtasks is even slimmer:

  • inherit (default; parent model)
  • claude-fable-5-thinking-high
  • composer-2.5
  • gpt-5.6-luna-high
  • gpt-5.6-sol-high
  • gpt-5.6-sol-xhigh

@Marcin_Miler How did you achieve that long list? I’m actually a bit envious.

Hey @Yishai_Elyada, if you belong to an organization that has restricted which models are available, those restrictions also apply to subagents.

Everyone starts from the same base list of subagent models, which is then filtered by what each account can access, including any models or variants your workspace admins may have disabled (for example, specific providers or the “fast” variants).

That’s why the list can look different from user to user.

Thanks @Colin

I’m pretty sure I can select grok-4.6 with all reasoning levels as a chat model, so I don’t think my org restricts it. But it’s not available as a subagent model. Chatting with Cursor it was pretty clear to me that this list is not controllable by anyone on our side (I also talked to our Cursor admin about it).

I sincerely think this is a bug, the list does not reflect what you say.

If you want I can send screenshots and chat transcripts to this effect.

Hey @Yishai_Elyada,

Two separate things stack here.

  • The base list that cloud subagents draw from currently includes only the “fast” variant of Grok 4.6. The regular reasoning variants you use in chat aren’t in it for anyone yet, which is a limitation on our side as discussed earlier in this thread.
  • Separately, organization model settings do also filter the subagent list (that part is controllable on your team’s side), and a restriction on “fast” variants would remove that one remaining Grok 4.6 entry, which is how you can end up with no Grok 4.6 for subagents at all even though chat works fine.

@Colin
Thanks, I see I missed that detail about the fast model. I think it makes sense now.

Yishai

P.S. As an aside, is there any authoritative place where I can search to see if there are open issues like this, and maybe subscribe to updates? The chat didn’t find this post (never mind the detail I missed), nor did searching the forum. I wrote my planning skills to adapt to the issue (it was constantly replacing grok-4.6 with GPT 5.6 Sol Extra High because the task tool couldn’t use the model, so I gave it a list of acceptable models for now) and it would be nice to get a push when it is resolved.

When there’s an update to this specific issue, I’ll update the thread!

Same limitation from the IDE Task tool, not only cloud subagents.

Use case: a project skill (/delegate) that takes the current Plan Mode plan and launches one background generalPurpose Task to implement it. I want that job on Cursor Grok 4.6 High, Fast off. The parent chat is often still in Plan Mode on fast mode, so inherit is not a workaround.