Agent overrode Task Models settings and billed Claude tokens instead of Cursor Grok
Steps to Reproduce
Configured Task Models → Explore Subagent Model = Cursor Grok 4.5 Medium.
Enabled primary model Cursor Grok 4.5 (Cursor models only; not Claude).
In Agent chat, prompted to implement the UnoQ “Profile-Driven Variant-Aware Menu Testing” plan and complete all to-dos.
The agent launched Task/subagents and explicitly forced Claude models (Claude Opus / Claude Sonnet), overriding the configured Cursor Grok Task defaults.
Tokens were billed against those Claude models instead of Cursor Grok.
After I pointed this out, later work used Grok, but the earlier Claude Task usage had already been charged.
Expected Behavior
Subagents/Tasks use my configured Cursor Grok models unless I ask otherwise
I checked the affected run. The earlier Claude calls were made by general-purpose Task subagents, not the Explore subagent. The Grok 4.5 parent explicitly selected Sonnet 5 and Opus 5 for those tasks.
When delegating work, the parent agent can choose the subagent type and model it considers the best fit for that specific task. The Explore subagent model setting only controls tasks launched specifically as Explore subagents. It does not apply to every general-purpose Task subagent, so your Explore setting was not overridden in this run.
The Claude calls consumed your included Other Models allowance, but they did not generate an on-demand charge. The Request ID you shared is from the later Grok-only portion of the session.
If you see an Explore subagent use a model other than the one configured for Explore, please share the Request ID from that affected parent turn and we can investigate it.
This is my request id which used Sonnet and opus. Thiughh it did not use on demand usage, it burned my api credits
ef564fba-672e-45b3-b7b3-f483f34eb0e3
Same issue class on Cursor desktop (Windows), Aug 10–11 2026. Facts only.
Selected model in picker: Grok / grok-4.5.
Local state.vscdb parent bubbles: modelInfo.modelName = grok-4.5 with Request IDs e.g. 9ad5f627-7e2a-4f43-a4c4-59259042faeb, a3dbe72f-ebd3-4612-8e00-040a09d5af72, 0336ed43-c5dc-40ad-978d-b5c857ddee9b.
Same DB also shows composer-2.5 / composer-2.5-fast in composerData:*, and the session transcript has multiple Task calls with explicit "model": "fast" (not the parent Grok slug).
In that Grok-labeled Agent session (history-chart / fake-landscape work): the agent claimed a “unified interface” while keeping three duplicate chart skeletons; cell-only layout hacks instead of shared layout; asserted satellite auto-tooltip / cell legend alignment that failed on device; long rewrite/revert cycles with little durable progress vs token spend.
