When running a bugbot /review my base chat was using Composer 2.5 (not fast). It spawned a few sub agents based on my response to the “Which review should run”. Within these chats the model appears to be Composer 2.5 (not fast), but on my dashboard (https://cursor.com/dashboard/usage) it appears as though they are being billed as fast. (See screenshot. Bottom entry of composer 2.5 appears to be my initial bug bot chat, then the spawned subagents are on fast based on timestamps and number of entries). Note my “Explore subagent mode” in settings is Composer 2.5 (not fast).
Steps to Reproduce
Start a chat on composer 2.5 and type /review.
When the which review to run prompt comes up select the third option and type “Both”.
Allow review to run and check dashboard for usage.
Oh I should also note I didn’t capture it but earlier when I instructed the chat to launch multiple subagents to fix independent bugs it launched the sub agents with GPT5.4 extra high but the base chat was using Composer 2.5.
Thanks so much for the detailed write-up and screenshots, this was really helpful.
You’re reading the situation correctly, and the short version is that nothing is being over-charged here. There are a couple of separate things happening:
The billing is accurate. The subagents spawned by /review really did run on Composer 2.5 Fast, so the Fast entries on your usage dashboard are correct. Composer 2.5 Fast is the default model for the /review subagents (the Bugbot and Security Review agents), so that part is expected.
The model label is what’s off. In the subagent chat (and in places on the usage view) the model can show as “Composer 2.5” without the “Fast” indicator, even though it’s actually running Composer 2.5 Fast. That label mismatch is a known issue on our side and something we’re actively working on. The underlying model and the billing are correct, it’s just the displayed name that isn’t surfacing “Fast”.
About the “Explore subagent model” setting. Good callout, but that setting only controls the Explore subagent specifically. It doesn’t change the model used by the /review subagents (which use Composer 2.5 Fast) or by general task subagents, so it’s expected that it didn’t change what you saw here.
On your follow-up about the GPT-5.4 subagents. When you ask the agent to spin up subagents to work on independent tasks, those general-purpose subagents can use whatever model the main agent picks for the job, which is why you saw GPT-5.4 subagents even though your base chat was on Composer 2.5. That’s also expected behavior rather than a billing issue. If you want to force disable Cursor from using GPT5.4 as a subagent, you can disable the model in settings and Cursor won’t be able to use it as a subagent. Otherwise Cursor will try to pick a good model for the subagent that best suits it for the task.
Hi Kevin, thanks for the detailed response. Is it possible to change the default model for the /review to be Composer 2.5 non-fast? (Or disable fast mode from settings as suggested for the GPT5.4 issue in bullet #4).
Can we get a setting or composer-2.5[fast=false] for built-in Bugbot/Security Review subagents (/review-bugbot, /review-security), same as custom subagents? On-demand usage shows large fast bills (~4M tokens per run) when doing local review before push, often with the parent on Auto. I also had Settings → Agents → Subagents → Composer 2.5, Fast OFF, but usage still billed as composer-2.5-fast — so those settings don’t seem to apply to review subagents (only Explore?). Any supported workaround, or is Fast fixed by design for IDE review?
