@condor I am curious: are you folks using the same Sonnet versions as Claude Code does? I am using both, and this issue here is impossible to replicate on CC, and on CC it gives you exact build version which I was never able to get with Cursor). Just curious if Anthropic is proxying calls differently based on where they are coming from.
they have said this is a bug, but I do not know why it would not be fixed by now
I’m not aware of who uses which precise version but will ask the team.
There were last week similar reports by CC users on Opus, but won’t go into details on those.
Personally I do not really know or have insight what Anthropic is doing internally or what system prompts they use in CC per model.
Thank you, makes sense. Just thought it is worth checking it, as it could have direct impact on what users see and report on Cursor side vs on CC directly.
My system prompt demands this info every request, I cancel and re-request if it answers as 3.5, the quality shows, night and day dif if it answers as 3.5 or 4.
I suspect they swap to 3.5 as a backup to mitigate high demand. Unsure if its Cursor or their provider doing this, but in any case if you request Claude 4 you ought to get it or be warned that they’ll give you a weaker model instead, a bad response in your code can be worse than no response.
1 Like
I’ve really tried to make Auto Mode work, but I’m done. Clean coding with GPT in Auto just doesn’t make sense – it keeps generating wrong code, wasting hours. Tasks that even free models handle correctly are consistently broken here. Sorry, but that’s it for me.
2 Likes
Unsubscribed because of this pricing ■■■■. Cursor is a good tool but pricing has to be fair, transparent and simple. Right now, the best I can do is unsubscribe.
3 Likes
It seems like the prices keep going up without ever stopping. I have multiple accounts, and I reach the limits faster and faster with use
1 Like
建议使用userRules限制一下，尽量节省地去使用token，并且每次应该开启新的聊天窗口，因为我发现大部分token都是被Cache Read消耗的
Can anyone help me? Cursor’s customer support won’t refund my payment to the original method. Instead, they gave me $144 in credit, which can only be used for future subscriptions and cannot be withdrawn. My emails to @hicursor have not been addressed, and my posts on the forum are either ignored or I’m told to email @hicursor again. I subscribed on April 13th and used the service until July 8th. I have been trying to process a refund from July 8th to July 28th, but I have still not been successful.
4 Likes
I was forced to switch to roo-code, using Orchestrator and different models for each mode. If Cursor could support similar features—allowing custom selection of thinking mode during planning, and choosing between GPT-4.1 or other auto models when coding—users would benefit greatly.
But Cursor will obviously prioritize strategies for revenue rather than user benefits
According to them, the price of cache is very low
The current pricing policy will kill Cursor, the pro rate is completely unprofitable and useless against competitor offerings. If before I had enough requests and even stayed until the end of the payment period, now I had to subscribe to another service to continue using neural networks. Goodbye Cursor!
At least $60
Will be switching to Claude code as well. Cache overload due to tool calls that trigger crazy token consumption make Cursor unusable except in auto (bad performances). Recent experiment by @mokhir show that Claude Code achieve same results with less API cost, and usage is reset every 5 hours, not every month.
More than anything, the most disappointing factor was the political-style communication from the team, presenting new plans as benefiting users, not answering most questions and cherry-picking messages to answer. Most of us would perfectly understand a token-based billing strategy, given a proper explanation as well as appropriate request processing that don’t cache 100x your input/output tokens. We also understand it’s not easy to develop such a solution, but you could at least
- communicate on this properly. You say you are working on this for next release, but what are the possible outcomes, when will we see it, what would be an ideal long-term billing strategy, etc?
- do not make users assume 100% of the new API cost overnight. If suddenly new models and agentic behavior make API cost explose, well, you are more than partially responsible. So the least we could expect is sharing cost before you find a solution, in order to slightly damage user experience instead of completely breaking it.
I used to advertise Cursor to friends and colleagues, and the few of them who tried it are also very disappointed with the current situation and switching to Claude Code.
4 Likes
I am not using cursor near as much as I used to but man, my dollar per output shot up tremendously, which is crazy considering how much I am using Claude Code now. they will say there is no problem, but there is
You are correct. I used to advertise Cursor alot to peers. Right now after they failed in even answering my support requests, I am moving on to Claude Code + 20 bux Cursor for the next 1 week. I am already on the 100 bux Claud Code plan and loving it so far. I wouldnt mind paying the 200 bux for Claude Code now.
So far 60% of my new codes are done by Claude Code and its working much much better than Cursor. And you are right, limits resets every 5 hours. Even hitting the limit just drops from Opus to Sonnet and I barely see the difference.
I guess this is me transitioning away. I spent close to USD500 on cursor this month. I dont mind paying cursor that one a monthly basis to be honest. (i have already informed my partners that monthly AI cost will hit 1k) But the screw up and non existence support now makes me irritated.
If after 1 week I see myself not using cursor at all, then i will just cancel my subscription.
Goodbye for now …
the cursor text input window is a huge help to right the CC terminal commands instead of having to use terminal LOL
never seen so much negative activity on the forums. I feel like I am flying through my usage even faster than last week. I am having to spread my 400 dollar ultra usage so thin to avoid getting hit with the insane fees. Grok 4 Heavy and 100 dollar Claude Code is a better and more affordable stack right now. This is so frustrating for a product that I really enjoy
2 Likes
Anthropic just introduced more rate limits
What’s changing:
Starting August 28, we’re introducing weekly usage limits alongside our existing 5-hour limits:
- Current: Usage limit that resets every 5 hours (no change)
- New: Overall weekly limit that resets every 7 days
- New: Claude Opus 4 weekly limit that resets every 7 days
- As we learn more about how developers use Claude Code, we may adjust usage limits to better serve our community.