I used up all of my API usage (Ultra) and once on-demand kicked in there was an immediate drop in quality outputs and the context windows would rapidly fill up from doing what appears to useless tool calling… and nothing ever seems to get done correctly any longer. I am normally on Opus 4.7 and GPT 5.5 and burnt by API limit in about 8 days so the drop in usefulness was dramatic and noticeable.
Opus spent ~$15 just thinking about how to not solve a problem and almost hit the context window in a couple minutes… Same with GPT 5.5… that is not what my experience had been while I still had API left.
Has anyone else notice a dramatic performance dropoff one it was on-demand?
Just want to be super clear: there’s nothing in Cursor that affects performance based on whether you’re using included or on-demand usage.
For Opus 4.7, I’d recommend paying close attention to your selected model effort level. Higher effort levels (xhigh, Max) can trigger a significant amount of thinking and tool calls. According to the Opus 4.7 docs, xhigh is targeted for:
Long-running agentic and coding tasks (over 30 minutes) with token budgets in the millions
That can quickly consume a substantial number of tokens. High and medium are worth a try!
I’ve noticed it also, but in my case it was less related to being in my on-demand usage and more related to the time of day.
During peak hours Opus started dropping context and making additional requests to do things that it could do without extra reasoning steps earlier in the day. This was working on a single task from a well laid out plan document.
Oddly it didn’t consume more of it’s context window according to the UI - but Cursor’s UI can’t be trusted when it comes to usage and billing, and neither can the usage dashboard. So who knows?
That was when I saw several line items in my usage dashboard where Opus showed “0” tokens.
The next day, the zeroes all turned into positive values and my on demand cap was met before I made any more requests. Billing basically said ‘it happens, its not real-time’