Hello.
I’ve been noticing this and reporting it to Cursor for several weeks now.
They are charging 2 credits for requests to the Sonnet 3.5 and 7 models, and 3 credits for the reasoning/thought model.
Is anyone else experiencing the same issue?
1 Like
Hey, the thinking model costs 2 requests, and this only applies to the Claude 3.7 Sonnet Thinking model. I don’t know where and how you could have gotten 3 requests. As proof, my screenshot is below, and yes, we announced this back in version 0.47.
Hello
See if this test works for you.
I also sent a video with the current credits, a thoughtful request, and final credits +3.
2 Likes
I have a video that I’m not going to post because it shows confidential information, but it proves how 3 credits are deducted.
So I’m not making anything up, and this is how the 500 credits disappear very quickly and unfairly.
1 Like
Here you go, more proof for you from right now.
Sonnet 3.5 - 1 credit
Sonnet 3.7 - 2 credits
Sonnet 3.7 - thinking 3 credits
1 Like
Make sure you dont have large context mode enabled as that will use more requests per prompt
Hello. Thank you very much for your reply.
Can you tell me how to activate/deactivate that option?
Thank you.
Failed requests are also charged. I don’t understand. Should it be Cursor’s fault?
100%, If a product charges you for something that doesn’t work or deliver usable results, that’s not a valid transaction. It’s the provider’s responsibility to ensure you get what you paid for.
Charging for failed outputs—especially when they’re unusable and reported repeatedly—can cross into deceptive or unfair billing. Once a company is made aware and doesn’t address it, that becomes negligence.
I’m actually experiencing the same thing at the moment.
Above was copy pasted before I submit the input.
Above is a screenshot of it escaping one block, making some code, and then entering a new block.
And then above you will see that I was charged a fast credit for that erronous output. So far every single output has been that error. I’m starting to experience like 98% errors, and 2% usable outputs. It will get into a zone and work for a bit and then just output intentional errors over and over.
Then above I ran into a series of responses that just copy the article back to me after it fails enough times charging me for an output when it’s not written any new content, only copied what I already wrote, word for word. It’s a simple request to convert a plain text document to markdown formatting and it fails almost 100% of the time.
Hey, can you check if you also have long context mode turned on?
My guess is with that and thinking, you get to 3 premium requests per prompt!
Hello.
Indeed, the deduction of that number of credits is due to the large context, something I wasn’t aware of.
Thank you all for your responses.
Best regards
1 Like