continue, but you keep switching to models with api costs not using auto why?
I was pinning Task subagents to claude-sonnet-5-thinking-high, which bills separately — I’ll leave model unset so they use Auto. Continuing the send-window work myself.
Hi @Steve_C, thanks for the second request ID. I traced it, and your main chat really was on Auto. The usage came from a subagent spawned during that same turn, a general task subagent running on Claude Sonnet 5 Thinking, which made roughly 45 model calls against the 3 on the Auto parent. One thing worth noting, nothing went to on-demand. Every call in that turn, parent and subagent alike, was billed as included usage on your Ultra plan, so what you saw move was your included quota rather than on-demand usage.
This part is expected behavior today, and it works the same way whether you pick Auto or a named model. A Cursor agent session can spin up subagents on whichever enabled model it judges best for the task, and the picker in Settings > Models > Task Models only governs the Explore subagent that handles initial codebase research. General task subagents, like the one that ran here, choose their own model at the moment they are created, from the set of models enabled on your account.
The hard guardrail is the model list itself. An agent cannot spawn a subagent on a model that is disabled in Settings > Models, so switching off Claude Sonnet 5 Thinking (and Claude Opus 5, which drove a similar block of usage overnight on the 26th) removes them from what the agent can pick across every subagent type. Two smaller things that help: custom subagents defined in .cursor/agents/*.md accept model: inherit, which keeps them on your session model (details at Subagents | Cursor Docs ), and a rule such as “subagents should use the same model as the main session” nudges the agent in the right direction, though rules guide rather than enforce.
my usage in other models went up like 20% from yesterday, and I watched it go up some today. Are you saying it even though it shows it burned it up it wont count?
The report says included, but it’s showing the other models in usage, not auto
Hi @Steve_C, it does count towards your API usage if a third-party subagent is spun up. It is still part of your included usage, so you aren’t incurring extra charges, but I understand that your API usage might be running out faster than you expect. The best way to prevent this is to (even temporarily) disable the third-party models you don’t want your Subagents to spin up in settings.
This is a design choice to help you get the best results from your agent’s queries - The Cursor Agent is allowed to spin up subagents of its choice that it thinks are the best fit for the job; however, the Agent is only able to spin up a subagent if it is allowed in your Cursor Settings. So if you disable all models except Cursor Grok, Composer 2.5, and Auto, it will only be able to use those models and won’t be able to spin up a model like Sonnet 5 as a subagent.
Hmm, interesting choice. I judicially manage my other model usage to get through the month. Disabling and re-enabling through the settings will be tedious. I often build a plan with Auto, then have Fable or another model review it.
I understand. Rules can help too. You can always create an always-applied rule like “ALWAYS USE SAME MODEL as PARENT FOR SUBAGENTs” but this is a behavior nudge rather than a hard-enforcement. We recommend the settings option for best enforcement.