So as we all know Cursor has max mode defaulted for various models (GPT 5.4, 5.5, Opus 4.7 etc). But 5.2 and 5.3 were not under this default.. until yesterday. I ran several requests under 5.2 high and xhigh and woke up this morning to over 400 used credits from my 1,000 just in a few prompts. Can someone confirm if they see this as well? Is this a bug or intended new behavior?
Hey, thanks for the report. To understand what exactly happened, and whether the requests really went out as Max Mode or if it’s a mismatch between the UI and the usage, we’ll need a bit more detail:
- The Request ID for a couple of the requests that used a lot of credits (Chat > three dots in the top-right > Copy Request ID).
- Your Cursor version (Cursor Settings > About) and your OS.
- A screenshot of the model picker in one of those chats. Can you see the Max Mode toggle there, and what state is it in for GPT-5.2 high / xhigh?
With that, we can check on our side how the requests were routed and whether Max Mode was applied.
Also, GPT-5.2 xhigh has historically been heavy on tokens for agent tasks because of the amount of reasoning, so some of the usage could come from that. But 400 credits for a few prompts looks suspicious, so let’s dig into it.
Hey,
One of the request ids: cb2e7c86-826d-4aff-a1eb-c7a2f42a2767
Version: V 2.4.31
To be clear:
- 5.2 and 5.3 have never been (and are still not) listed under MAX mode.
- The issue started last night. Prior to that, 1 credit would be deducted.
- 400 requests was exaggerated. I consumed just under 300 in 7 prompts.
- I’m now being helped by Sam from Cursor who requested that his colleague look into it.
Prior to yesterday, this was the standard display
Update: the issue seems to be resolved now. Not sure if Sam or someone internally had triggered the fix but its back to normal.
Cheers
ReqId: 70be3d06-6dc8-4241-9f9e-55a5fe39aae5
scary right , i remember this bug , it was also solved while the ticket person is still confused about where the problem is and that the numbers look normal to them
Thanks for the update and for the detailed screenshots, they really help. The usage history does show a regression: the same gpt-5.2-xhigh / gpt-5.3-codex-xhigh models that were billed as 1 credit per request before Apr 28 around 23:00 UTC, and again after the morning of Apr 29, suddenly started being billed as MAX mode (by tokens) in the window between those times, even though they aren’t labeled MAX Only in the model picker. That isn’t intended. On the legacy request-based plan, these models should be 1 credit per request.
Good news that you already have a thread open with Sam. Credit reconciliation for that period (around 300 credits) will go through that, so no need to duplicate it.
If you notice it happening again, send a new Request ID here and we’ll take a look right away.