The pricing has been confusing in many ways thus far. It is improving, but still confusing as of yet.
Cursor doesn’t impose a strict limit. They guarantee “at least” $20 of usage, but they may potentially offer more than that. Hence why you are seeing a “bonus” of $36.14. Exactly how that is determined I cannot say, and we may never know.
I’ve become an extreme user of Cursor’s agent, and I’ve recently moved to the Ultra plan. I started on Pro for the first few months, was on Pro+ for about a month, where I was given usage to around $100 or so, then also paid over $150 of pay-as-you-go usage. So I was given some extra, but in the long run, I ended up spending more than I would have if I had just gone strait to the Ultra plan for $200/mo and the usage limits on that are WAY beyond Pro+.
Usage isn’t determined….well, deterministically, I guess is the best way to put it. I’ve run out of claude-4-sonnet usage on some days, then on later days I was able to use it again without having to pay-as-you-go. Some of it is based on the overall usage of any given model at the current moment. If you select a specific model, say claude-4-sonnet or gpt-5-fast, it is a lot easier to burn through your allotment of requests for your given plan.
If you use the Auto mode of Cursor, it distributes your requests over all the models it is capable of using, and according to their docs you can get more usage out of any given plan by using Auto. When I use Auto, it usually uses claude-4-sonnet (there are just certain output characteristics, no thinking steps, etc.), but Auto can choose to use any model that Cursor deems fit, and sometimes you will notice that the output changes, you might see thinking steps appear (or disappear), etc. Sometimes the nature of the code will change as Auto switches between models for different requests.
Auto will usually allot you a lot more usage each month, than sticking with a single model. Claude usually allows for a lot fewer requests than other models (not sure how GPT-5 will pan out there in the long run), Gemini seems to allow for about 3x or so what is allowed with Claude. For mundane stuff, I’ll often explicitly use GPT-4(something), as it allows for significantly more requests, and I’d rather not burn Sonnet requests on simple asks, research, etc.
You can manage your usage each month. Cursor’s recommended approach is Auto, but sometimes that isn’t as effective as it could be (especially now, there seems to be a severe bug with Auto usage right now clamping conversation lengths to nearly zero). When you are using a specific model, try to use the one best suited for the task. You can enable Custom Modes, and configure various modes to use specific models and allow or disallow certain behaviors (i.e. I have a Research custom mode that uses GPT and disallows code file editing, and a Planning custom mode that uses Gemini and disallows code file editing, etc.) Each mode can be assigned a hotkey, so you can quickly switch modes.
Also, I think the latest version of Cursor finally properly supports using a different model in each tab (previously, and I didn’t realize this at first, when you switched the model, that was GLOBAL, and affected ALL TABS! Oh, did that lead to a few mishaps!)
Anyway. There are a number of ways to optimize your Cursor plan usage. I’d delve into the docs and explore the UI and learn how to optimize cursor for your own use cases.