I reached 100% usage. Previously, I couldn’t access Cursor AI at all, but now I can still use it after reaching the limit and work with slow model.
Is this new feature or what?
Hey, this is expected behavior, not a bug. On Free, after you use up your included usage, you can still access Auto, but responses go through a slower queued pool. Cursor says this in the banner: “You’ve reached your limit. Responses may be slower.”
On the usage page, these requests are marked as Free and don’t count toward billing. Speed depends on load, so it can drop a lot during peak hours. If you need consistent speed and a higher limit, Pro or Pro+ give faster responses and more usage.
Is this a new thing? Because I know it wasn’t the case for me last month, so I just wanted this clarification.
This is part of how Free works right now. After you use up your included usage, you can still access Auto through a slow pool. Speed depends on load, so during peak hours requests might not go through at all. It looks like you hit one of those times last month, so it felt like access was completely blocked.
Does this apply to credits usage as well?
I’m on the Pro plan, and I’ve used up 100% of my included usage, and now using credits. I’ve found that it’s still slow, even though I’m not using Auto, and I’m able to use a specific model. The usage shows up as Free, but credits are deducted as expected.
It’d be nice if I could use my credits as regular requests instead of slow requests.
Do you mean On-Demand when you say credit usage? If on demand is on, you can use any model you want with cost of on demand model cost. Otherwise, if you’re on free slow mode, you will see a banner saying so.
On-demand is disabled.
No banner saying I’m on slow requests.
Credits balance is being used. Usage still shows as free - because of credits I guess?
I did briefly switch to GLM5.2 before switching back to GPT-5.5, and observed slowness only when using GLM5.2. Could have been a one-off too.
But I’m not sure what/how to test how the slow rule is applied to credit usage, or if I am in fact switched to slow usage.
It’s all good, though. I’m just recording this here in case someone else runs into this 
Hey, based on the screenshots everything lines up, it’s not a bug. The reason is that On-Demand Spending is disabled (you can see it in the first screenshot: “On-demand spending is currently disabled”).
How it works:
- Your included API usage is at 100%, and on-demand is off. So requests to models after the limit go through as a free fallback. That’s why the table shows Type “Free” and Cost “Free”. These are not paid requests, and no credits are used.
- Your credit balance ($15) is applied automatically only to on-demand (usage-based) charges. Since on-demand is off, there are no charges, so no credits get deducted.
If you want requests to be normal paid requests and have your credit balance used, enable On-Demand Spending in your billing settings. After that, named-model requests (like gpt-5.5-medium) will be billed at the on-demand cost, and your credits will be applied automatically.
Let me know if anything behaves differently after you turn it on.
I don’t think the part about “credit balance … is applied automatically only to on-demand (usage-based) charges” is correct.
From your reply in How to use the credits obtained by cursor from Refer friends - #7 by deanrie (bold styling is mine)
…referral credits are usage credits. They’re automatically applied to usage-based charges like premium requests and overages above your plan’s included limit. You don’t need to activate anything manually or use a promo code.
The $15 is the credits I’ve used, and it keeps increasing the more I use Cursor. The credit balance is $675.
This—
If you want requests to be normal paid requests and have your credit balance used, enable On-Demand Spending in your billing settings. After that, named-model requests (like gpt-5.5-medium) will be billed at the on-demand cost, and your credits will be applied automatically.
… doesn’t seem to be the case either. Named-model requests are using credits just fine, without having to enable on-demand spending.
But I agree this isn’t a bug, and it’s working as it’s supposed to. Was just curious about slowness when I tried GLM5.2… and that’s a non-issue, because GPT5.5 seems to be working at usual speeds now.
Summarizing what I’ve observed so far with respect to referral credits, just to document it:
- You do not need to enable on-demand usage in order to use credits balance
- Credits balance is automatically deducted after you reach 100% of the included usage of your plan, if you use named-model requests
- The usage against credits is shown as “Free” in the usage reports
- Cursor usage that uses credits doesn’t seem to be routed to slow requests (not tested, just purely based on my observation, atleast with GPT5.5)
Please feel free to correct me if I’m wrong here.
Thanks for the careful breakdown. You’re right, and my post wasn’t accurate. I’ll fix it:
- The “$15” in your case is used credits, not what’s left. Your balance ($675) goes down as you use them, so that’s correct.
- Usage credits, including referral credits, apply to usage above your included limit automatically. You don’t need to enable On-Demand Spending for that. Named-model requests can be billed from credits, which matches what you’re seeing.
- In usage reports this shows up as “Free” because there’s no extra charge beyond the subscription. It’s covered by credits.
Your final summary is correct, so I’ll leave it as a reference point for others.
About the slowdown on GLM5.2 while GPT5.5 is fine, it sounds like a one-off load issue or something specific to that model. If it keeps happening, send the Request ID and we’ll take a closer look.
Update: I’ve been daily-driving the GLM5.2 for a few days now, and everything works well