Cursor is using Composer 2.5 Fast as subagents. It drains my tokens pool.
My IDE is configured to use Composer 2.5, not Composer 2.5 Fast for initial research.
Steps to Reproduce
It’s happening when I’m spinning more challenging task.
Hey there!
This is a known bug on our side, not your setup. Your main agent is correctly using the model you pick (GPT-5.6 Sol Medium). The issue is only the built-in Explore subagent: when your main model is a pricier one like GPT-5.6 Sol, it currently falls back to Composer 2.5 Fast even with Fast toggled off. That Fast usage is genuine exploration work (real subagent runs), not a metering error - it’s just running the pricier variant than you intended, which is what’s eating your tokens.
Two ways to avoid the Fast spend while we sort out the fix:
Cursor Settings → Models → Explore Subagent Model → Disable — stops the subagent entirely, so your main agent handles exploration itself.
The Fast fallback only kicks in under a pricier main model. For exploration-heavy work, running Composer 2.5 (Fast off) as your main model makes the Explore subagent correctly use non-Fast Composer 2.5.
I’d avoid Inherit from parent here - with GPT-5.6 Sol as your main model, the subagent would then use that even-pricier model.
This is something we’ve seen reported, so I’ve logged your report with the team and will update here when there’s progress. More on how the Explore subagent works here.
so let me get this right. I use gpt 5.6 because it’s more powerful and efficient. so efficient in fact it knows that it can delegate some work to smaller models like composer 2.5, but cursor still charges me the gpt 5.6 pricing for all the work the composer is doing ? is that what’s going on? I have an ultra plan and have an on demand monthly limit of $200 that always gets me through the month. I ran out of on demand credits in about 10 days and then I added another $100 to the budget and went through that in about an hour. What is Cursor doing about crediting people back the money we have spent on this bug?
@Chris_Craft - On the billing part first, because that’s the crux.
You’re not charged GPT-5.6 pricing for the work the subagent does. Every model is billed at its own token rate, so when your GPT-5.6 agent hands exploration off to a Composer subagent, that subagent’s tokens are billed at Composer rates, which are actually cheaper per token than GPT-5.6, not more.
The real problem is narrower. There’s a known bug where the built-in Explore subagent runs the Fast flavor of Composer 2.5 instead of the standard one, and Fast costs more per token than standard Composer 2.5. It’s a real bug on our side, not your setup and not a metering error. It’s just the wrong variant running.
Two workarounds that actually stop the spend right now:
Cursor Settings → Agents → Subagents, and set the Explore subagent to Disabled. Your main agent then does exploration itself, so there’s no Fast subagent spend.
For exploration-heavy work, run Composer 2.5 (Fast off) as your main model. Under a Composer main model, the Explore subagent correctly resolves to standard (non-Fast) Composer 2.5.
I’d avoid “Inherit from parent” here, since with GPT-5.6 as the main model the subagent would just use that pricier model instead.
On credits: I don’t want you eating on-demand cost for a bug on our end. Since that’s tied to your account and billing, email us at [email protected] from your account and mention this Explore-subagent Fast issue, and the team can review your specific on-demand usage and look into a credit for the bug-attributable portion. That’s not something I can action from the forum.
This is a known issue we’ve had reported, and I’ve logged your report with it. More on how the Explore subagent works here.
