Selecting Grok 4.6 still bills Claude Fable 5.1 / Other Models; in-flight Fable continues after picker switch

Support Bug Reports
, , ,
1

Where does the bug appear (feature/product)?

Cursor IDE

Describe the Bug

On Pro, I mainly use Cursor Grok 4.6 (Cursor Models pool). Other Models still climbed to 27% this cycle while Cursor Models is only 5%.

Usage dashboard (UTC) for 2026-09-15 shows two included events billed as claude-fable-5-1-thinking-high:

  • 11:08 AM — 890k tokens
  • 11:09 AM — 10.294M tokens

That is ~11.2M Fable tokens in one morning session. Local renderer logs for the same conversation (50e2dedb-0919-4a4b-953c-b60ce790f8f3) show:

  • 04:08 PDT: picker actually had Fable selected (buildRequestedModel idSource=selectedModels[0], selectedModelIds=claude-fable-5-1, matchingSelectedModel=true, maxMode=true, 300k context, thinking-high)
  • 04:11 PDT: same conversation picker switched to grok-4.6
  • a worktree window kept running Fable on that conversation until ~04:26 PDT

After switching to Grok, Other Models still rose. The 7-day Usage chart also lists gpt-5.6-sol-max even though I did not intend to use it.

This matches known reports that Task/Cloud helper subagents do not inherit Grok and bill Other Models (threads 165737, 170233). Please also check whether an in-flight Fable turn / worktree keeps billing Fable after the picker is changed to Grok.

Steps to Reproduce

  1. Open Cursor IDE Agent chat on Windows (Pro plan, on-demand disabled).
  2. Start a long Agent task. In this case the picker resolved to claude-fable-5-1-thinking-high with 300k context (conversation 50e2dedb-0919-4a4b-953c-b60ce790f8f3 at 2026-09-15 04:08 PDT).
  3. A worktree window opens for the same conversation and continues the agent loop.
  4. Change the model picker on the parent chat to Cursor Grok 4.6 (logged at 04:11 PDT).
  5. Do not stop the already-running agent.
  6. Check Dashboard → Spending: Other Models percentage rises. Dashboard → Usage shows claude-fable-5-1-thinking-high events (11:08–11:09 UTC, 89万 + 1029.4万 tokens).
  7. Continue working only with Grok 4.6 in later chats; Other Models remains much higher than Cursor Models (27% vs 5% this cycle).

Expected Behavior

If Grok 4.6 is selected, usage should come from the Cursor Models pool. Changing the picker to Grok should stop Fable billing on that conversation, including worktree/subagent windows. Helpers should inherit the parent model or stay in the same pool. Settings → Models restrictions should apply to subagents.

Operating System

Windows 10/11

Version Information

Cursor 3.20.21 (Windows 10/11). Logged in hooks as cursor_version 3.20.21.

For AI issues: which model did you use?

Intended: Cursor Grok 4.6 (cursor-grok-4.6-high / high-fast). Also billed: claude-fable-5-1-thinking-high. Chart also shows gpt-5.6-sol-max.

For AI issues: add Request ID with privacy disabled

Conversation 50e2dedb-0919-4a4b-953c-b60ce790f8f3. Generation 86670005-d82a-48ad-ac38-967ac983a038 (Fable). Later 7af9d47e-8af2-42af-bb35-410f01206ab0. Grok chat 6d71825f-910e-4aec-bcf1-b7c9d3b0c30f.

Additional Information

Plan: Pro $20/mo. On-demand disabled. Spending this cycle: Cursor Models 5% used, Other Models 27% used, Grok Bot weekly 14%. Related threads: Agent Task subagents spawn on Fable/GPT without user requesting those models and Cloud Agents subagents use different models than selected . Please review Other Models charges from the Fable 5.1 events on 2026-09-15.

Does this stop you from using Cursor

No - Cursor works, but with this issue

7

Hey @bz_liu, thanks for the detailed report, and for including the conversation and generation IDs. That let us reconstruct exactly what happened.

Grok was never billed as Fable. Both Fable entries on the dashboard are the run you started at 04:08 PDT, when the picker was still on Fable 5.1 (Max, 300k). Changing the picker to Grok at 04:11 only applies to the next message you send. A turn that is already running keeps the model it started with until it finishes or you press Stop, and that is true in every window.

The part that should not have happened is the second window. When the worktree opened in an Agents window, that window marked the conversation as “in progress” and automatically resumed it, so the same turn ran twice. Your IDE stream was stopped at 04:11 when you switched to Grok, but the Agents window kept running the Fable copy until you stopped it at 04:27, and that copy is the 10.29M-token entry. We are tracking this and I have added your report.

Two things to try in the meantime:

  1. In the Agents window, open Settings, then General, and turn off “Continue Interrupted Agents”.
  2. If you open an Agents window or a worktree window while an agent is still running in the IDE, check that window for a second running copy and stop it there.

On gpt-5.6-sol-max: that came from a Grok chat on Sep 9 at 12:29 UTC that spawned three helper agents, and the helpers ran on gpt-5.6-sol-max instead of inheriting Grok. That is a separate known issue we are already tracking (helpers not inheriting the parent model), and I have added your report there too.