I have seen this posted before, without resolution from the Cursor team, so I have disabled privacy mode and have provided a request ID. I believe the team are aware of this issue, so am providing more insight.
When asking Claude 3.7 Sonnet in Agent mode to add an edit modal to my existing Vue component, it hangs and just “stops”
I’m unsure how to make it occur in a new project, but I get it consistently within my current project for the same prompt and file. Changing the prompt makes no difference. Starting a new chat does not resolve it either. I have only included 1 Vue file in the initial request.
This is stopping me from progressing with Cursor within my project, as it refuses to continue make changes, yet continues to use my credits per failed request. Does not occur with Claude 3.5 Sonnet. I am paying for Pro.
Model: Claude 3.7 Sonnet
Cursor version: 0.46.8 (Universal)
OS: M2 Max Sequoia (15.3.1)
Request ID: 8636bdd1-459e-450c-96ab-7f90dd26e2bc
Hi, I’m looking into this. The specific error seems to be that sonnet 3.7 is not doing as well a job adhering to max output token limits for tool calls compared to 3.5. Looking into fixing this.
I’m pretty sure this should not have counted as a fast request usage, I see the log Marking request as errored for usage because of error which should mean that it was not counted as a fast request.
If you’re seeing different let me know and I’ll also investigate the credit usage.
Hi Zackh1998 - thanks for the prompt response. I was monitoring the “Recent Usage Events” log from the Cursor website, perhaps I was misinterpreting it. I saw an initial premium request when the initial prompt was sent off, and was unsure if this was being reversed when I encountered the “stopped” message. I tried this a few times with the failure, and saw additions in the log each time, but wasn’t paying attention to the actual request counter / 500.
I’ve managed a workaround for now by not including any pill files in the initial prompt, and have asked Claude to break the workload out into smaller chunks. This seems to avoid running into the max output token limit quickly. Thanks for looking into this and hope to see a fix soon.
Another engineer is looking into the tool call errors themselves. I looked into the fast request usage, and it looks like it should not be charged when it errors.
It looks like the website doesn’t show any indication of that in the request logs, so I’ve update the website page to make it clearer, like in the image below:
let me know if you don’t see the “not charged for” messages, and a screenshot of the requests, and I can get someone to refund some fast request usages.
yes i’m also wasting my credits for sure… working with an old 0.45 version now… but my code got messed up since the latest update… lot’s of unclosed { } 's and such. Are to keep the logic of your code also in the AI momentum. As it needs to learning again from scratch what you are doing…
sorry for butting in here. there are so many bug reports with 0.46.x cursor issues and i have not received a single response. so yeah we do get charged for bad broken cursor features. like attaching terminal lines to chat shows them but doesnt pass them to AI. and a lot of other issues like not sending requests to AI api at all (but pro subscription checks pass successfully) etc.
Where can I check if they are really being spent? On the web I see that they are spent, I have spent about 30 fast requests since it began to fail on Friday.
I just posted this as a new thread because I didn’t see this current one, however errors indeed are consuming credits, here is a screenshot with both the counter and the 3 errors in a row
Some users above have reported that even though the log shows “not charged for” - credits are still being consumed regardless (see aoji’s screenshots above).
Please can you confirm if this is the case? If so, can you also get someone to look into refunding fast request usages? I’d like to be assured that credits are absolutely not going to waste due to system errors out of our control.
Hi any credits being charged from errored requests should have been fixed Tuesday Night. Let me know if you’re still seeing any credits going up from errored requests (they might temporarily go up during the request, but should go back down after the request errored).
If you’re still seeing them increase from errored requests, please send me screenshots and I’ll investigate.
I’m having this issue now - every edit claude tries to make times out now. I tried using a fresh chat and also restarting cursor. Literally nothing works (on the latest version, im pro). Fun times.
I’m having the same issue—Agent and Claude Sonnet 3.7 attempt to generate a response, but then I get ‘Stopped.’ In my account, it shows as ‘Included in Pro ✕ (but errored, so not charged for),’ yet the request counter still increases. I’ve just wasted a few dozen requests dealing with this. This has happened before, so I may have wasted many more premium requests.
Just to be sure, I tried a few more times—and yes, it ‘stops,’ says ‘so not charged for,’ yet the premium request counter increases with every failed attempt.
